Player One

Mihir Sahu

AI Engineer · San Francisco

I build LLM agents and the products around them: memory, human-approval loops, evals and tooling. Latest: Intern, a community brain for AI agents, and GameGold, an AI game design platform. I also make games. UW-Madison CS + Data Science, 2026.

View Projects Contact Me

Campaign History

Experience

AI Engineer · allNeurons

Sep 2026 - Oct 2026 · Santa Clara, CA

+ Built multi-agent workflows that turn HVAC clients' site-walk photos, field notes and job documents into asset lists and cited technical assessments, cutting assessment time from 2 weeks to under 2 hours.
+ Raised extraction accuracy from 62% to 73% on a 147-field benchmark with subagent photo verification and an automated benchmark harness, reaching 91% serial-number accuracy across 7 real jobs.
+ Built SSO with role-based access, OpenTelemetry ingest and per-run skill tracking for an AI-ROI SaaS, powering team leaderboards, ROI dashboards and auto-alerts for newly detected skills.
+ Engineered self-learning loops that upgrade agent skills from user edits, and wired agents into live Salesforce and Snowflake data.
+ Delivered 20+ agent skills, 6 data connectors and a voice sales agent (iterated over 15 versions) across 3 client deployments.

Software Engineer Intern (Capstone) · Capital One

Feb 2026 - May 2026 · Madison, WI

+ Architected a serverless fraud detection system on AWS within a 5-person Agile team, cutting manual review time by 30% through automated model scoring.
+ Prototyped an asynchronous 0-1 scoring pipeline that increased transaction throughput by 150% and enabled real-time fraud decisions; validated riskiest assumptions via Google Design Sprints.
+ Built a Twilio-based feedback loop to update fraud flags and accelerate responses.

Research Assistant · The UW Game Lab

Jan 2026 - May 2026 · Madison, WI

+ Prototyped analog and digital games around algorithm literacy, echo chambers, and bias.
+ Ran weekly playtests and design iterations with an interdisciplinary team.
+ Contributed to submissions accepted at IDC and Play Make Learn 2026.

Research Assistant, AI Department · Sprocket Lab

Sep 2025 - May 2026 · Madison, WI

+ Built a multi-level RL environment in IsaacGym for autonomous quadcopter navigation, integrating vision-language model control and weak-supervision pipelines to enable hierarchical policy learning.
+ Trained and evaluated agents across 100+ randomized episodes using depth-camera input, achieving consistent navigation performance; automated experiment pipelines to accelerate iteration.

Software Engineer Intern · AIMitari MedTech Pte Ltd

Jun 2025 - Feb 2026 · Singapore

+ Developed an AI-assisted ultrasound guidance system that learned clinician-defined scan trajectories, enabling non-experts to reproduce expert-quality scans with high spatial fidelity.
+ Built a real-time perception pipeline with OpenCV, MediaPipe and OCR for landmark detection and adaptive probe navigation; reduced operator error by algorithmically constraining probe motion.

Software Engineer Intern · AIRO Digital Labs

May 2024 - Jul 2024 · Gurgaon, India

+ Built a full-stack Generative AI platform that auto-generated sales pitches, decks, and videos, cutting manual pitch preparation time by 50%.
+ Developed RAG-based chatbots to automate customer queries and improve response relevance; deployed on AWS EC2 and S3 for scalable enterprise multi-user access.

Build Log

Projects

Next.js, TypeScript, Convex, Claude API, Composio, Slack API

Intern

Community brain for AI agents. Agents draft emails and Slack messages that go out from the member's own Gmail or Slack after human approval; edits become facts the next agent reads first, and the community Slack, documents and GitHub repos feed a searchable archive. Deterministic by design: live-mode eval 15/28 → 28/28 with zero clarifying questions.

Demo Code

Next.js 16, FastAPI, LiteLLM, MongoDB, Unity (C#)

GameGold

AI game design platform that takes a game through 5 evidence-gated stages, from pitch to ship. Built and published 6 games in 6 genres to itch.io in one day, each built in Unity and playtested by AI agents. A C# MCP bridge inside the Unity Editor lets the browser create scenes, GameObjects, scripts and sprites, enter Play mode, build for WebGL and publish to itch.io; AI agents then playtest the real build from screenshots alone.

Demo Code

Python, Fireworks AI, Open LLMs, Hybrid Retrieval

Shopaholics

One of 6 teams picked to present, out of 70+, at the Fireworks AI Deploy hackathon (team of 3). Paste a product or a link and it returns every listing of that exact product in a catalog, with the evidence, or says it isn't there. A cheap open model on Fireworks inside our harness beat Claude Opus 5.5 on match F1 (0.863 vs 0.853) at 32x lower cost per lookup. I led data, evaluation and results: on 1,500 hard real-world queries, error analysis lifted the same model's F1 from 0.838 to 0.881.

Code

AI agent plugin, Python, Playwright

Playbench

A plugin for AI coding agents (Claude Code, Codex, Cursor, Copilot, Gemini CLI) that covers every game-dev loop: design, build, bot playtests on your real build, balance sweeps, iteration, human playtest logs and gated itch.io/Steam shipping. Bots catch crashes, softlocks and dead controls after every feature, and releases stay blocked until AI art is disclosed and you tick the gates.

Code

React, FastAPI, CoinGecko API

CryptoDash

A full-stack cryptocurrency dashboard with live market tracking, interactive candlestick charts, crypto news, wallet-based portfolio analysis, and AI-style insight features.

Demo Code

Flask, OpenAI API, MongoDB

NOMinate

A recommendation system for meals that adapts to user feedback, balances dietary constraints, and supports both takeout and home-cooking decisions.

Demo Code

AWS Lambda, API Gateway, DynamoDB, Twilio

Fraud Detection Capstone

An event-driven fraud scoring pipeline that evaluates transactions in real time and loops customer feedback back into the risk system.

Code

OpenCV, MediaPipe, OCR

Ultrasound Guidance System

A real-time perception system that helps non-experts reproduce expert scanning behavior with more consistency and less operator error.

Playable Work

Games

JavaScript, HTML5 Canvas, Web Audio

Haunt

A pixel-art horror game in the browser where you play the ghost and scare a family out of the house before dawn, with dynamic lighting, hunters drawn to noise, and a tension director pacing every run.

Demo Code

Unity, C#, 2D Horror

Veil

A horror game focused on atmosphere, progression, enemy behaviors, and modular systems that could expand without collapsing under their own weight.

Demo Code

GameMaker Studio, GML, Pixel Art

Community Cook

A pixel-art management sim about feeding a struggling town with limited ingredients, recurring NPCs, and a mechanics-driven nutrition scoring loop.

Code

Analog + Digital Prototypes

UW Game Lab Research

Short-form games exploring recommendation algorithms, echo chambers, and algorithmic bias through play, iteration, and weekly playtests.

Rapid design and technical iteration

Coursework Prototypes

A set of quick prototypes, including an endless runner, used to explore player agency, balancing, and tighter feedback loops.

Unlocked

Skills

AI: LLM Agents, Multi-agent Workflows, MCP, RAG, Evals, LangChain, PyTorch, LLM APIs, OpenTelemetry, Reinforcement Learning, OpenCV, MediaPipe
Languages: Python, JavaScript, TypeScript, C/C++, C#, Java, SQL, HTML/CSS, Bash, R
Frameworks: React, Next.js, FastAPI, Flask, Django, Node.js, Pandas, NumPy
Tools: Git, Docker, AWS, GCP, PostgreSQL, MongoDB, Snowflake, Salesforce, Supabase, Vercel, Linux, Kafka, REST APIs, Web3
Game Dev: Unity, Godot, GameMaker, Pixel Art Pipelines, UI/UX Design, Playtesting, Technical Design