Kartikey Agrawal

Products Built and shipped

1. HostelMate

   Live demo online (still being polished); preparing for its first college

   Indian college hostels still run on paper registers, WhatsApp complaint groups and handwritten leave forms.

   One platform for students, wardens and parents: attendance, leave, complaints, mess, fees, visitors and lost-and-found.

   • Attendance you can't mark for a friend: a rotating HMAC-signed QR code, checked together with a 100 m GPS fence.
   • Face check in the browser (blink, motion, five angles), so verifying a student costs nothing per scan.
   • Complaints triaged by an LLM (Llama 3.1 on Groq): category, urgency and a summary for the warden; a 30-day view spots recurring problems.
   • Fees paid online through Razorpay, every payment's signature verified; visitors go through warden approval.
   • Live notifications over Redis and Socket.io instead of polling.
   • 260+ Jest tests in 23 suites and 59 Playwright end-to-end tests, run by a pre-commit hook and CI; an 80% line-coverage floor.

   

   

   

   Sole engineer (1,124 commits), 2026 – nowNext.js, Express, Supabase (PostgreSQL), Redis, Socket.io, Groq, Razorpay, Docker, TurborepoLive demoSourceDocs

2. StockSense

   In development

   Retail investors in India see a P/E of 24 and can't tell if that's healthy, because what's normal differs by sector.

   Stock research that scores every NSE company on six fundamentals against its own sector, and explains each score in plain language.

   • Live prices for 2,200+ NSE stocks, with no API calls while the market is closed.
   • Reads 400-page IPO prospectuses and pulls out the risks in under 30 seconds.
   • Screener, IPO scoring, a geopolitical risk map and paper trading.

   

   

   Built the product, data pipeline and interface, 2026 – nowNext.js, Python, FastAPI, Supabase, pgvector, FinBERTSource

3. FoodBridge

   Finished

   Restaurants throw away good food while NGOs nearby run short.

   A marketplace where donors list surplus food and NGOs or volunteers claim it before it expires.

   • Live countdowns and urgency colours on every listing.
   • Claims are atomic, so two NGOs can never take the same food.
   • Tracks food rescued, meals provided and CO₂ prevented.

   

   Solo, Harvard CS50x final project, 2026Python, Flask, SQLiteWatch the demoSource

Research district Papers, benchmarks and tools

1. Do small models talk themselves out of right answers?

   Accepted at ICTCS 2026, presenting mid-December

   A common test truncates a model's reasoning and forces an answer. If an early forced answer is right and the final one wrong, people call it 'overthinking'. But forcing the answer is itself an intervention.

   A two-tier protocol that separates what the probe forces from what the model actually commits to, run on DeepSeek-R1-Distill-Qwen-1.5B, Qwen2.5-3B and Llama-3.2-3B over GSM8K and MATH-500.

   • 210 uninterrupted traces; 11 flagged as degradation (5.2%).
   • Two annotators, the second blind, agreed on all 11 (Cohen's κ = 1.00).
   • Only 1 of the 11 is genuine: the probe measures itself as much as the model.

   Co-author, with Mayanka Gupta, Kiran A Gupta, Utkarsh Mandal, Abhigyan Shekhar and Parth Bhat, 2026PyTorch, Hugging Face, GSM8K, MATH-500

   11 of 210 traces looked like a model abandoning a right answer.

   • 6 never declared an answer before the token budget ran out
   • 3 were wrong without ever declaring a correct answer first
   • 1 was still correct once the probe was removed
   • 1 genuinely gave up a correct answer for a wrong one

2. Gearbox

   Almost complete, paper in progress

   Expensive AI agents spend most of their tokens on small, routine subtasks a cheaper model could do, if anyone could trust the cheaper model's 'done'.

   An automatic transmission for AI agents: the agent hands subtasks to cheaper models, keeps working, and accepts an answer only after it passes a real test. Failures move up to a stronger model with exactly what failed.

   • Cheap-first with checking: 95.7% on HumanEval+, against 92.7% for always using the 27B model, at about a third of the compute.
   • The expensive model wrote about half as many tokens.
   • Tested on 542 problems and 12 open models on a lab GPU; our grader matches published scores within about 1 point.
   • Plugs into Claude Code, Cursor and Antigravity as an MCP tool.

   Built the tool and ran the benchmarks, 2026Python, LiteLLM, MCP, OllamaSource

   Accuracy on HumanEval+

   Always use the 27B model

   92.7%

   Cheap first, check, move up on failure

   95.7%

   Compute cost (relative, lower is better)

   Always use the 27B model

   27

   Cheap first, check, move up on failure

   8–9

3. Punarvaani (पुनर्वाणी)

   Code complete and tested on CPU; GPU training next

   People who lose their voice to throat or oral cancer surgery, a stroke or ALS can still move their lips, but no sound comes out. India has one of the highest rates of oral and throat cancer.

   Reads a person's silent lip movements through a phone camera and speaks for them in Hindi, in their own voice rebuilt from old voice notes and videos: lips to text (an LLM fixes look-alike sounds) to cloned speech, or lips straight to speech units.

   • Every stage is written and tested on CPU (80+ tests): lip crop, a Hindi audio-video data engine, an AV-HuBERT encoder with LoRA, both speech paths, voice rebuilding, real-time streaming and evaluation.
   • On a synthetic lip-reading world, per-user personalisation cut the character error rate from 92% to 15.5% with 64 clips.
   • Next: mine Hindi video, record volunteers mouthing silently, and train on GPUs.

   Built the codebase, 2026 – nowPython, PyTorch, AV-HuBERT, LoRA, MediaPipe, IndicF5Source

4. MetaGuard

   In progress

   AI tutors can help so much that people stop learning.

   Psychology × AI research on how people learn with AI tutors.

   • Details after the study is published.

   Researcher, 2026 – nowLearning science, LLMs

5. Amazon ML Challenge: one company, many spellings

   Top 5%

   Match the same business across three messy sources with about 12.5 million records: typos, swapped legal forms, missing addresses, and names written in Devanagari, Tamil, Telugu or Gurmukhi.

   A three-stage pipeline: cheap candidate search on rare name and address keys, a LightGBM matcher with about 70 features, and a decision layer that enforces one-to-one matches.

   • A transliteration dictionary learned only from the training pairs, so Indic-script names meet their English versions.
   • Score 0.965 out of 1, rank about 1,800 out of 35,000 teams.

   Team DayTrippers, four members, Sep 2026Python, LightGBM, rapidfuzz, sparse_dot_topn

   विजन प्रोडक्ट्स प्राइवेट लिमिटेड=Vision Products Private Limited

   1. Find candidates on rare name and address keys
   2. Score each pair with about 70 features
   3. Keep one match per record

Market street Smaller builds

About And the training ground