Papers, essays and case studies by Mitansh Mittal, a triple major in Finance, Computer Science and Economics at Santa Clara University.https://mitanshm.com/en-ushttps://mitanshm.com/research/cant-hallucinate-can-still-be-wrong/https://mitanshm.com/research/cant-hallucinate-can-still-be-wrong/Typed-decision models answer in fixed-shape fields, such as a yes or no, one option from a list, or a score, in a single forward pass, and they are sold as unable to hallucinate. I tested what that guarantee leaves out. BEMA is an open, from-scratch reproduction of the interface: one small transformer encoder with binary, 77-way and regression heads, trained only on human-labeled data and calibrated with temperature, Platt, isotonic and split-conformal methods. On clean held-out text it looks trustworthy: temperature scaling cuts the binary head’s expected calibration error from 0.0159 to 0.0046. Then I added two adjacent-character typos per query, the ordinary fast-typing kind. Accuracy on the 77-way head fell from 81.9% to 51.2% (n = 3,080), while mean confidence fell only from 0.791 to 0.612. The same gap appears on a second, 151-way dataset. Retraining on typo-augmented data halves the accuracy drop, and the gain carries over to typo types it never trained on, but it does not close the gap. The output guarantee is real: the model cannot emit an answer outside its schema. It can still be confidently wrong on input squarely inside its own domain. A noise-robustness number belongs next to ECE whenever a model’s confidence is sold as a safety signal.Wed, 23 Sep 2026 00:00:00 GMTPapercalibrationrobustnessconformal predictiontyped decisionsevaluationhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/research/disagreement-as-signal/https://mitanshm.com/research/disagreement-as-signal/A language model is most dangerous when it is confidently wrong, and its own confidence is a weak guard. This position paper describes an architecture I built nine times between April and August 2026. Agents do the work, deterministic checks that no model can vote on test it, a council of models from other labs challenges it, and a human takes every split decision. From those builds come three working theories: adjudication beats synthesis, the model that writes should not check its own work, and agreement is not proof. Only the third has data behind it so far. In a two-model cross-check of 94 form-field mappings, all 3 known errors sat inside the 7 disagreements, though the agreements were never audited. On 20 planted citations, a string matcher and an LLM judge agreed on 19 and were both wrong on 2 of those, so their failures overlapped instead of cancelling. Recent work finds that errors correlate across model providers too. I therefore treat independence as a property to measure, not a feature to buy, and propose an evaluation on three tasks with checkable ground truth: error rates when checkers agree and when they disagree, adjudication against synthesis, and self-review against cross-lab review.Wed, 23 Sep 2026 00:00:00 GMTPapermulti-agent systemsLLM evaluationLLM-as-judgehuman-in-the-loopposition paperhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/research/killing-good-ideas/https://mitanshm.com/research/killing-good-ideas/Most of what I start, I stop. This essay lays out the kill rules I actually use and the builds that taught them to me: cutting a feature I could not build honestly, throwing out results that looked too good, keeping the engine when the product goes quiet, and a lesson I learned backwards, that the market map belongs before the code.Wed, 23 Sep 2026 00:00:00 GMTEssayjudgmentproduct strategyresearch practicehello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/research/uncorrelated-failure-modes/https://mitanshm.com/research/uncorrelated-failure-modes/The design rule behind most of what I build: never trust one source of truth. Independent methods vote, deterministic code does the arithmetic and keeps the books, every fact carries its provenance, and a person decides whenever the checks disagree. This essay explains the rule through my own systems, including where it has already failed me: checks that were supposed to be independent, and weren’t.Wed, 23 Sep 2026 00:00:00 GMTEssaydesign principlesreliabilityhuman-in-the-loopAI systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/research/sts-generative-ai/https://mitanshm.com/research/sts-generative-ai/An explainer on how large language models work and why they fail. They are trained to predict the next token, not to check the truth, so fluent output can still be wrong. The paper defines hallucination through NIST AI 600-1’s “confabulation” and walks through documented failures, from fabricated citations to the sanctions in Mata v. Avianca. It covers how training data shapes errors and lays out a four-step routine for verifying model output. Then it turns to my own systems. The multi-agent and council designs I build supply three working theories: adjudication beats synthesis, the writer should not be the checker, and agreement is not proof. BEMA, the model reproduction I architected and ran by directing AI coding agents, supplies a measured result: two adjacent-character typos cut a classifier’s accuracy from 81.9% to 51.2% (n = 3,080) while its mean confidence fell only from 0.79 to 0.61. All eight references are real and checkable.Tue, 22 Sep 2026 00:00:00 GMTAcademic work · STS (Science, Technology and Society)generative AIhallucinationAI policycourseworkhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/ai-game-pipeline/https://mitanshm.com/work/ai-game-pipeline/Generating and play-testing genre games on Roblox with coding agents: three genres in parallel, one routed to verify.Tue, 01 Sep 2026 00:00:00 GMTCase studyexploredAI SystemsGames & SimulationLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/bema/https://mitanshm.com/work/bema/Stress-tested a “can’t-hallucinate” model design: 2 typos cut accuracy 82% → 51%, confidence only 0.79 → 0.61.Tue, 01 Sep 2026 00:00:00 GMTCase studyresearchResearchAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/blacksite-null/https://mitanshm.com/work/blacksite-null/Co-op extraction-horror shooter for Roblox: 173 strict Luau modules from parallel AI agents, gated by written contracts.Tue, 01 Sep 2026 00:00:00 GMTCase studyin-buildGames & SimulationAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/collapse/https://mitanshm.com/work/collapse/24-player falling-floor party game on a heap-scheduled tile engine where idle tiles cost nothing per tick.Tue, 01 Sep 2026 00:00:00 GMTCase studyarchivedGames & SimulationLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/exit-liquidity/https://mitanshm.com/work/exit-liquidity/An incremental game about running pump-and-dumps, on a deterministic order-book sim that counts every victim.Tue, 01 Sep 2026 00:00:00 GMTCase studyin-buildGames & SimulationFintech & RegTechhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/jobapp/https://mitanshm.com/work/jobapp/Local-first job-search OS: finds roles, tailors résumés behind a truthfulness gate, fills forms, waits for your OK.Tue, 01 Sep 2026 00:00:00 GMTCase studyin-buildAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/night-day/https://mitanshm.com/work/night-day/A Brave fork with two persistent personas, Night and Day, and a one-key switch, compiled clean from source.Tue, 01 Sep 2026 00:00:00 GMTCase studyexploredLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/the-trials/https://mitanshm.com/work/the-trials/24-player last-one-standing Roblox game with server-validated combat and a headless two-match smoke harness.Tue, 01 Sep 2026 00:00:00 GMTCase studyarchivedGames & SimulationLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/chispen/https://mitanshm.com/work/chispen/Compiles a written process into a playbook that runs, measures itself and proposes its own fixes.Sat, 01 Aug 2026 00:00:00 GMTCase studybuiltAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/accountward/https://mitanshm.com/work/accountward/Turns a conservator’s bank statements into a balanced draft California court accounting, reconciled to the penny.Wed, 01 Jul 2026 00:00:00 GMTCase studyin-buildFintech & RegTechAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/ai-diligence-market-map/https://mitanshm.com/work/ai-diligence-market-map/Teardown of the AI due-diligence market: six segments, feature-by-feature defensibility, and what actually survives.Wed, 01 Jul 2026 00:00:00 GMTCase studyexploredLabFintech & RegTechhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/cleanplay/https://mitanshm.com/work/cleanplay/Collaboration: built with Fluxxion88. Graph-based device trust scoring that targets gold-selling rings in online games but never penalizes a shared family IP.Wed, 01 Jul 2026 00:00:00 GMTCase studybuiltCollaborationsAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/concord/https://mitanshm.com/work/concord/Multi-vendor M&A diligence engine: 11 agent desks, cross-lab challengers and a model-free tie-out on every quote.Wed, 01 Jul 2026 00:00:00 GMTCase studyarchivedFintech & RegTechAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/intern/https://mitanshm.com/work/intern/Collaboration: built with Fluxxion88. An LLM loop that turns one hand-made example into a deterministic script with no model inside.Wed, 01 Jul 2026 00:00:00 GMTCase studybuiltCollaborationsHackathonsAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/model-routing/https://mitanshm.com/work/model-routing/Routed models by task, cost and vendor across three apps; the routing worth keeping is the rule for who checks the work.Wed, 01 Jul 2026 00:00:00 GMTCase studyexploredLabAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/quorum/https://mitanshm.com/work/quorum/Local-first private-AI diligence: open-weight models via Ollama, five workstream leads, a risk committee, an IC memo.Wed, 01 Jul 2026 00:00:00 GMTCase studyarchivedLabAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/redl/https://mitanshm.com/work/redl/A local-first AI workbench: model store, chat, a gated coding agent and a multi-agent Council in one Tauri app.Wed, 01 Jul 2026 00:00:00 GMTCase studybuiltAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/visual-hand-track/https://mitanshm.com/work/visual-hand-track/Collaboration: built with Fluxxion88. Webcam gesture recognizer: finger-state heuristics over MediaPipe hand landmarks drive two-hand gesture actions.Wed, 01 Jul 2026 00:00:00 GMTCase studybuiltCollaborationsLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/warrant-forge/https://mitanshm.com/work/warrant-forge/Collaboration: built with Fluxxion88. Estate-settlement engine where no fact reaches a decision without a verbatim quote found in a source document.Wed, 01 Jul 2026 00:00:00 GMTCase studybuiltFintech & RegTechHackathonsCollaborationshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/warrant-portal/https://mitanshm.com/work/warrant-portal/Kotlin Android app that turns an estate engine’s blocked decisions into plain-language questions for the executor.Wed, 01 Jul 2026 00:00:00 GMTCase studybuiltMobileHackathonshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/bluecollarpal/https://mitanshm.com/work/bluecollarpal/Procurement and virtual-card control plane for trade contractors: text a request, approve it, pay with a one-time card.Mon, 01 Jun 2026 00:00:00 GMTCase studyin-buildFintech & RegTechMobileAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/colossus-wake/https://mitanshm.com/work/colossus-wake/One-night Godot 4.6 action-platformer prototype set on a sleeping titan, built by directing OpenAI’s Codex agent.Mon, 01 Jun 2026 00:00:00 GMTCase studyexploredGames & SimulationLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/forgecouncil/https://mitanshm.com/work/forgecouncil/The council as an AI engineering team: reviewer agents plus a judge, in a safe-by-default CLI and desktop app.Mon, 01 Jun 2026 00:00:00 GMTCase studyexploredAI SystemsLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/habit-tracker-bot/https://mitanshm.com/work/habit-tracker-bot/Collaboration: built with Fluxxion88. Telegram habit tracker with per-user ownership checks, repository tests, CI on every push and Railway deploys.Mon, 01 Jun 2026 00:00:00 GMTCase studybuiltCollaborationshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/job-board-aggregator/https://mitanshm.com/work/job-board-aggregator/Collaboration: built with Fluxxion88. Daily ingester for Hacker News “Who is hiring?” threads with constraint-enforced dedup and a paginated FastAPI API.Mon, 01 Jun 2026 00:00:00 GMTCase studyin-buildCollaborationshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/research/econ-3-agricultural-subsidies-debate/https://mitanshm.com/research/econ-3-agricultural-subsidies-debate/Closing speaker for the CON side of a team debate on the motion that “American subsidies on agricultural products are harming the poorest people in the world and should be immediately eliminated.” A closer can’t introduce new evidence, only reframe it, so I held the motion to its exact wording. Our evidence showed real damage from U.S. farm support: jobs lost to the sugar program, deadweight loss, and 4 billion dollars in WTO-authorized EU tariffs over the Boeing dispute. That damage lands on U.S. industries and rich trading blocs, not on the world’s poorest. My closing reframed the case around incidence. Most least-developed countries are net food importers, so cheaper world food helps them, and immediate elimination would hit them with a price shock. I also separated crop insurance and disaster aid from the subsidies that drive overproduction.Mon, 01 Jun 2026 00:00:00 GMTAcademic work · ECON 3economicstrade policydebatecourseworkhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/up-to/https://mitanshm.com/work/up-to/A “less planning, more hanging” plans app: one-tap invites, a drag-to-paint availability grid, an iMessage stub.Mon, 01 Jun 2026 00:00:00 GMTCase studyexploredMobileLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/research/ctw-2-architecting-the-social-fabric/https://mitanshm.com/research/ctw-2-architecting-the-social-fabric/A literature review that synthesizes three sources into one argument: cohesion in a diverse society is built deliberately, through spaces and processes, and does not follow from proximity. Patel’s case for civic “builders” sets the frame. Eck’s distinction between diversity, a demographic fact, and pluralism, active engagement, sharpens it. A 2025 mixed-methods study of the Jordan Downs public-housing redevelopment in Watts, Los Angeles, grounds it: 21 stakeholder interviews and a survey of 647 residents found higher social cohesion there than at two comparison sites. The review closes with implications for planners and policymakers. I then built it into a single-page HTML and CSS research poster sized for a 48 × 36 inch board, with a source-synthesis table, a conceptual framework diagram and a print stylesheet.Sun, 31 May 2026 00:00:00 GMTAcademic work · CTW 2urban policypluralismliterature reviewcourseworkhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/research/omis-40-daily-step-count-study/https://mitanshm.com/research/omis-40-daily-step-count-study/A survey study with a 7-person team: what predicts how much Santa Clara students walk? We surveyed 29 students on their self-reported daily steps and eight candidate drivers: distance from campus, class load, attendance, transportation, exercise frequency, social outings, job status and year. Then we worked the conditionals. Students who exercise at least three times a week cleared 8,000 steps 60% of the time (9 of 15), and students who go out at least four days a week did so 55% of the time (5 of 9). We closed on the limitations: a sample under 30, a convenience sample, self-report bias, outliers and missing confounders.Sun, 31 May 2026 00:00:00 GMTAcademic work · OMIS 40 (Statistics & Data Analysis I)statisticsconditional probabilitysurvey designcourseworkhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/adde/https://mitanshm.com/work/adde/Cross-vendor M&A diligence: per-provider agent swarms whose findings a council of provider representatives arbitrates.Fri, 01 May 2026 00:00:00 GMTCase studyexploredAI SystemsFintech & RegTechLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/oblivion/https://mitanshm.com/work/oblivion/A privacy-first messenger and wallet prototype: Waku E2EE messaging, an Ethereum wallet, one encrypted local vault.Fri, 01 May 2026 00:00:00 GMTCase studyexploredLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/cli-task-tracker/https://mitanshm.com/work/cli-task-tracker/Collaboration: built with Fluxxion88. Terminal task manager with sections, JSON persistence and a per-task stopwatch that banks time on Ctrl+C.Wed, 01 Apr 2026 00:00:00 GMTCase studybuiltCollaborationsLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/nexus/https://mitanshm.com/work/nexus/Solo-built wargame adjudicator: five Claude agents debate while a 10,000-run Monte Carlo cascade does the math.Wed, 01 Apr 2026 00:00:00 GMTCase studyarchivedAI SystemsGames & SimulationHackathonshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/project-omni/https://mitanshm.com/work/project-omni/The native Rust half of a hold-to-talk AI dictation app: global hotkey, streamed audio, screen capture, typed-back text.Wed, 01 Apr 2026 00:00:00 GMTCase studyexploredAI SystemsLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/rezonyr/https://mitanshm.com/work/rezonyr/Collaboration: built with Fluxxion88. AI sprint planner: RAG over a team’s docs, Drive and Discord → five weekly bets → team vote → per-person plan.Wed, 01 Apr 2026 00:00:00 GMTCase studybuiltCollaborationsAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/the-council/https://mitanshm.com/work/the-council/Agents do the work, other models check it. Built nine ways, then measured, with the negative results reported.Wed, 01 Apr 2026 00:00:00 GMTCase studyresearchAI SystemsResearchhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/econode/https://mitanshm.com/work/econode/Security-first FastAPI + Next.js scaffold for a crop-genomics screening idea, shelved the same day.Sun, 01 Feb 2026 00:00:00 GMTCase studyarchivedLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/eyeos/https://mitanshm.com/work/eyeos/Webcam eye tracking as an input device: MediaPipe gaze moves the Windows cursor behind a four-gate safety model.Sun, 01 Feb 2026 00:00:00 GMTCase studyexploredLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/fitfindr/https://mitanshm.com/work/fitfindr/Android app that names your outfit from one photo, wired to a 4B vision-language model built to run on the phone’s NPU.Sun, 01 Feb 2026 00:00:00 GMTCase studybuiltMobileAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/meetwise/https://mitanshm.com/work/meetwise/Electron meeting copilot: live transcript and screen context feed one-line suggestions to an always-on-top overlay.Sun, 01 Feb 2026 00:00:00 GMTCase studybuiltAI SystemsLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/notes-app/https://mitanshm.com/work/notes-app/Offline speech-to-notes in Java: on-device Vosk transcription, Gemini-formatted notes and screenshot Q&A.Sun, 01 Feb 2026 00:00:00 GMTCase studyexploredLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/riptide/https://mitanshm.com/work/riptide/Put a live AirSim drone behind my team’s ocean-cleanup console: Flask RPC bridge, cameras, LiDAR and missions.Sun, 01 Feb 2026 00:00:00 GMTCase studybuiltHackathonsGames & SimulationMobilehello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/esp32-lora-bench/https://mitanshm.com/work/esp32-lora-bench/ESP32-S3 and LoRa bring-up: three boards incl. a LilyGO T-Deck, Meshtastic 2.7 staged, my own I²C diagnostic.Thu, 01 Jan 2026 00:00:00 GMTCase studyexploredHardware & MakingLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/slugbites/https://mitanshm.com/work/slugbites/Drone-delivery ordering for UC Santa Cruz dining halls on live-scraped menus. 2 awards at CruzHacks 2026.Thu, 01 Jan 2026 00:00:00 GMTCase studybuiltHackathonsMobileGames & Simulationhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/linkleap/https://mitanshm.com/work/linkleap/My Nov 2025 pitch deck for an auto-apply job agent for new grads, later rebuilt properly as JobApp.Sat, 01 Nov 2025 00:00:00 GMTCase studyexploredLabhello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/aerobites/https://mitanshm.com/work/aerobites/V1 of my campus drone-delivery idea: a React ordering app plus an AWS Bedrock agent back end, written as IaC.Wed, 01 Oct 2025 00:00:00 GMTCase studyarchivedHackathonsAI Systemshello@mitanshm.com (Mitansh Mittal)https://mitanshm.com/work/cad-3d-printing/https://mitanshm.com/work/cad-3d-printing/Scaled models to the build limits of three printers and ported projects across printers and slicers, TPU included.Thu, 01 May 2025 00:00:00 GMTCase studyexploredHardware & Makinghello@mitanshm.com (Mitansh Mittal)