Sairam Krishnan
Projects

Ondoway

A local storyteller in your pocket that adapts as you walk.

Co-founder & Chief Technology OfficerAI travel companionPrivate betaConcurrent with Apple

I own product strategy, AI architecture, and end-to-end engineering in the founding team. I designed, architected, coded, and operate the whole system.

Walking tours rarely know where a listener is, what they care about, or how much time they have.

A good tour ties accurate stories to the exact corner in front of you, your interests, and your walking constraints.

A provenance-aware Neo4j knowledge graph grounds every story.

A deterministic tour engine plans and sequences the route, and GPS triggers narration at each stop.

FastAPI backend, authoring tools, and a Flutter mobile app.

Private beta with early testers, including Wharton Executive MBA peers, in New York City and Paris, the first launch city.

I turned their feedback into route-sequencing priorities.

Every spoken sentence traces back to a sourced beat.

I evaluate with frozen human-labeled cases, blinded comparisons, evidence checks, and code-enforced gates.

I rejected a cheaper LLM judge after calibration exposed false rejections of valid paraphrases, separating model-quality failures from truncation before making the cost decision.

I instrument model and prompt identity, tokens, latency, responses, evidence, and the cost of every generation and narration, failed attempts included, to guide debugging, model selection, and unit economics.

A travel companion that adapts as you walk.

Persuaider

Practice, simulate, and evaluate the hard conversation.

Creator & AI Engineering LeadAI negotiation simulation and coachingIndependent work

I built the complete product: agent behavior, multiagent simulation, evaluation, and its multi-provider AI operating model.

Negotiation training is hard to personalize, repeat, and evaluate consistently.

Practice needs credible competing interests and feedback grounded in what was actually said and agreed.

Scenario authoring and in-character AI personas with distinct motivations, confidential objectives, and walk-away limits.

Interactive practice and stateful AI-versus-AI negotiations.

LLMs extract the negotiated terms, then code scores utility, deal feasibility, walk-away, and value left on the table against fixed rules.

Production-grade application with persisted transcripts and usage, security controls, transparent model failover, and a deep automated test suite.

I score negotiation economics separately from conversational skill.

Coaching follows configurable, weighted, framework-based rubrics, and every quotation and number in it is checked against the transcript in code, returned for bounded repair, and withheld if it still fails.

Side A, private objectivesSide B, private objectivesLLM deal extractionDeal scored in codeCoaching checked against the transcript

Gestiq

Deliberate desktop actions from hand motion.

CreatorOn-device computer visionIndependent work

I designed and built the native macOS product, extending my Carnegie Mellon capstone: a 2048 puzzle game played with hand gestures.

Camera input is uncertain.

Useful gesture control must tell an intentional action from jitter, tracking loss, and the hand returning to rest.

Apple Vision hand-pose estimation runs on device.

Temporal recognition normalizes motion to palm size, which makes two-hand controls possible and routes commands by the active application or website.

Working macOS application with a SwiftUI interface and menu-bar controls.

Video never leaves the Mac.

Confidence, timing, and target identity travel through the whole perception-to-action path.

The app processes only the latest frame, rechecks the target at the final dispatch, and shows live diagnostics, so uncertain camera input never fires a desktop action.

Lore / DependIQ

The verification layer AI coding agents lack.

I designed the product, the graph model, the algorithms, and the senior-engineer audit.

Repository-local tooling cannot explain the organization-wide blast radius of a dependency change.

PostgreSQL holds identity and history, Neo4j traverses dependencies, and task-aware routing sends each job to a hosted or local model.

Working multi-tenant product with more than 200 tests.

AI-assisted engineering needs explicit dependency truth before it needs another generator.

One changeBlast radius

Crucible

Engineering workflows that make AI agents show their work.

I turn engineering judgment into reusable skills, agent perspectives, hooks, adversarial review, and executable acceptance rules.

An agent’s confident report is weak evidence that a change is correct, complete, or recoverable, and prompt-only guidance is easy for an agent to ignore when delivery pressure rises.

Twelve skills and four pattern guides built on structured agent reports, claim-level evidence checks, risk-based verification, mutation and undo evidence, concurrent review, deterministic gates, and bounded revision loops.

Open-source toolkit with runtime hooks and regression tests, and the working counterpart to my management research.

Separately, my internal agent-engineering work supports daily operations and was adopted and extended by QA, Operations, and upstream and downstream teams.

Reliable agents come from executable constraints.

I keep model judgment separate from the code that checks evidence structure, enforces workflow decisions, and controls when work can proceed.

Model judgmentAgent reportCode checks itClaim by claimWork proceeds

More I’ve built

Shipped to strangers

Claude Code Profile Switcher

2026

A VS Code extension that switches Claude Code between accounts, such as work and personal. It moves the ~/.claude link, sets CLAUDE_CONFIG_DIR, and shows the active profile in the status bar. Developers have installed it more than 4,400 times.

VS Code Marketplace

Synonym graph

2015-16

synonymscrawler maps words by degrees of separation through a thesaurus, and randomwordgenerator serves random words; both have been on PyPI since September 2016, and an outside developer contributed a fix I merged. A Node app draws the crawl as a D3 tree you expand word by word.

GitHub

Parallelax

2015

An npm package that runs several Blockspring functions in parallel and merges their results into one object. It ran my stock-briefing and dictionary APIs side by side, and was downloaded more than 1,500 times in 2016 and 2017.

GitHub

Built with AI agents

Peeraxis

2026

A harness that runs coding agents for me. I approve one feature card; it writes a hidden test, builds the feature, has a model from a different company check the work, and records a demo I accept or reject. It delivered its first Persuaider feature with no message from me, and it built 12 of its own 43 commits, including its own approval screens.

GitHub

Newsleaf

2026

I lead product and technology with co-founder Katera Mujadidi. An eight-step pipeline reads six news sources, extracts events with LLMs, merges duplicate stories, records when each event happened and when it was learned, and shows which way each story moves crude, LNG, tankers, and airlines. Pre-launch.

Live site

Taught

The Bias Files

2025

A ten-page interactive comic, drawn entirely in HTML and CSS, that teaches four kinds of AI bias (historical, representation, aggregation, and evaluation) through real cases, including Amazon's scrapped hiring model and the Gender Shades study.

Live site

The AI ROI quest

2026

The reader becomes Chief Innovation Officer of the Kingdom of TechRealm and works through the costs and benefits of an AI investment, one discounted-cash-flow decision at a time.

Live site

Built to learn

Machine learning by hand

2016

I worked through Coursera's Machine Learning exercises in Octave and wrote each algorithm from its equations: linear regression by gradient descent and the normal equation, regularized logistic regression, a one-vs-all digit classifier, backpropagation for a 400-25-10 network, and learning curves to tune regularization. Around it, statistics in R and functional programming in Scala.

GitHub

Hand-2048

2014

My Carnegie Mellon capstone team played 2048 with hand motions in the air. A C++ tracker on a depth sensor sent each swipe over TCP to a Node proxy that the browser game polled; a dead zone and re-arming made one swipe exactly one move. Gestiq grew from it.

GitHub

One survey app, three ways

2013-14

I built the same survey product three times: Spring MVC with Hibernate and MySQL, Spring MVC on Google App Engine, and Django on App Engine. Login, image uploads, transactional saves, salted password hashing, and survey suggestions based on what others took.

GitHub

MedHub

2014

An Android forum for medical questions, built with Meghna Thakur. Patients post symptoms, doctors and patients reply, and replies are voted up or down, with separate doctor and patient accounts and about 3,800 lines of layered Java.

GitHub