KRISHAN SAINI · PRINCIPAL AI ENGINEER · REMOTE, INDIA (UTC+5:30) · US & EU HOURS

I design, build, and run AI systems. Personally.

12+ years shipping production systems. The last three building AI-native products solo: retrieval over 19.4M court judgments, HIPAA clinical voice, autonomous outbound. When you hire me, I write the code, in your repo, through your review process.

See the systems reachout@krishansaini.com

DIRECT CLIENT BUILDS · EMBEDDED IN YOUR TEAM · FRACTIONAL CTO · PRINCIPAL ROLES

NOT TECHNICAL? START HERE →

SYS.01 · COURTMESH.AI

CourtMesh.ai

Lawyers find the controlling judgment in seconds, and every answer shows its source.

AI legal research for Indian litigators: semantic search and citation-grounded chat over 19.4M Supreme Court and High Court judgments.

19,400,000SC & HC judgments indexed
213,540statutory sections
49,487Acts
CACHINGrepeat queries skip inference entirely
STATUS · LIVE ↗courtmesh.ai, in production

bar widths · log scale

PYTHON · NODE.JS · REACT/TS · OPENSEARCH · KUBERNETES

OPEN COURTMESH.AI · TRY A SEARCH ↗

Plain-English version: a live product you can open right now: the same end-to-end job I’d do for you.

CourtMesh.ai architecture Pipeline: parse, chunk, embed, hybrid search on OpenSearch, rerank, citation-grounded chat, running on autoscaled Kubernetes workers. K8S · AUTOSCALED EMBEDDING & INFERENCE WORKERS PARSE CHUNK EMBED OPENSEARCHHYBRID RERANK CITECHAT

Select any stage: the engineering decision behind it.

fig. 01 · the production architecture, simplified.

CourtMesh research interface: a semantic search over case law with court and year filters, matched judgments, and an AI summary with citations
from courtmesh.ai · open it live above

I designed the retrieval pipeline end to end (parsing, chunking, embeddings, hybrid vector-plus-metadata search on OpenSearch, reranking) and built citation-grounded chat so the model can only answer from retrieved judgments, never fabricate case law.

Response caching cuts LLM cost and tail latency. The system runs on Kubernetes with autoscaling embedding and inference workers. It also ships GA4/UTM growth instrumentation and a marketing analytics console.

DEEP-DIVE · how the citation grounding works →

SYS.02 · ONCALL.AI

OnCall.ai

Doctors stop typing notes, and nothing enters the record without a physician’s signature.

A clinical voice product built to HIPAA Security Rule controls: transcribes, summarises, and medically codes physician–patient phone calls.

ASR→CODEone pipeline, call to codes
SIGNEDevery output clinician-reviewed
HIPAAencryption, audit logs, access control
STATUS · PRIVATEclinical deployment · live walkthrough on a call

PYTHON · WHISPER ASR · REACT/TS · LLM PIPELINE

Plain-English version: a product where mistakes matter, built so a doctor stays in charge.

OnCall.ai architecture Call audio flows through ASR then LLM summarisation and medical coding inside a sealed PHI boundary; clinician review and sign-off sit outside the seal as the human in the loop. CALLAUDIO ASR LLM SUMMARISEMEDICAL CODE CLINICIANREVIEW SIGN-OFFHUMAN AUTOMATED PIPELINE · PHI ENCRYPTED · AUDIT-LOGGED · ACCESS-CONTROLLED

Select any stage: the engineering decision behind it.

fig. 02 · the production architecture, simplified.

An ASR-plus-LLM pipeline tuned for clinical accuracy, with strict PHI boundaries between services.

Because no model should sign a medical record, clinicians review, edit, and sign off on every output in a React dashboard I built for exactly that. The human step was designed first, not bolted on.

SYS.03 · AUTO_GTM

AUTO_GTM

Outbound that researches, writes, and follows up on its own, with a hard gate before anything sends.

An autonomous outbound engine: sourcing, enrichment, LLM scoring, sequenced sends, and replies. Every message passes one chokepoint.

GATEDauto-replies, never instant
HELDsensitive threads wait for a human
BLINDscore-before-reveal lead judging
STATUS · IN USE DAILYruns my own outbound; I am the customer

PYTHON · NODE.JS · LLM SCORING · EMAIL INFRA

Plain-English version: I trust it to send email under my own name.

AUTO_GTM architecture Pipeline: source, enrich, LLM score behind a gate valve, sequence, reply gate, with an amber human-hold branch for sensitive threads. SOURCE ENRICH LLM SCORESCORE-BEFORE-REVEAL SEQUENCE REPLY GATE HUMAN HOLDSENSITIVE THREADS

Select any stage: the engineering decision behind it.

fig. 03 · the production architecture, simplified.

It sources and enriches leads, scores them with an LLM behind a score-before-reveal guard, and runs the sequenced sends itself.

Replies are answered automatically, but gated: sensitive threads are held for a person; nothing goes out until someone clears it. Most of the engineering here is restraint, deciding what the system may not do alone.

APPENDIX A · Job Search Engineturns messy job postings into clean, searchable listings with alerts · LLM-NORMALISED · REACT · NODE · MONGODB

EVERY SYSTEM ABOVE · DESIGNED AND BUILT SOLO, END TO END. CLIENT WORK IS THE OPPOSITE SHAPE: YOUR REPO, YOUR REVIEW PROCESS, YOUR TEAM.

TRACE · 2014 → NOW · SPAN: 12Y+

Principal Engineer, AI ProductsINDEPENDENT · VIA MY OWN ENTITY · 2023–PRESENT

Built CourtMesh.ai, OnCall.ai, and AUTO_GTM: retrieval, inference, infrastructure, and product. Client engagements via the Toptal network include Gartner, a cloud re-architecture that cut monthly infrastructure cost ~20%.

Principal DevOps EngineerPIXIS (AI AD-TECH) · 2021–2023

Model-serving infrastructure on GKE, Pub/Sub async pipelines, petabyte-scale BigQuery. Monolith to microservices; CI/CD with canaries and auto-rollback; shipped to production daily.

Lead DevOps & Cloud ArchitectBOOKING HOLDINGS · 2018–2021

GKE clusters serving millions of travel searches a day.

AWS → GCP migration~22% OPEX CUT
median incident response~40% FASTER
Senior Cloud Consultant, DevOps & SREVMWARE · 2016–2018

Zero-downtime SaaS migrations to VMware’s GCP hybrid cloud. Disaster recovery that was actually tested.

Software EngineerEARLIER ROLES · 2014–2016

REST APIs, real-time tracking, routing, payments: the years when I learned that shipped beats clever.

◆ 2014 · BE COMPUTER SCIENCE, PANJAB UNIVERSITY, CHANDIGARH

A decade running production infrastructure is why the AI products ship, and stay up. The Booking Holdings numbers above happened on my watch, not near it.

INVENTORY · CAPABILITIES

The tools change. The judgement is knowing which to use when.

AI / LLM SYSTEMS

LLM integration: OpenAI, Anthropic · citation grounding · prompt engineering & evals · agents & tool use · prompt caching, LangChain, Whisper ASR

RETRIEVAL & DATA

RAG & hybrid retrieval, embeddings & chunking, reranking, vector search (OpenSearch, pgvector, Pinecone) · Postgres, Redis, MongoDB, BigQuery · Pub/Sub, Pulsar

BACKEND

Python, TypeScript / JavaScript, SQL, Bash, FastAPI, Node / Express, REST & GraphQL, microservices, async workers

FRONTEND

React, Next.js, Tailwind, design systems, browser extensions

CLOUD & INFRA

Kubernetes · GCP (GKE, Cloud Run, Vertex AI) · AWS (EKS, ECS, Lambda, S3, DynamoDB) · Terraform, Ansible, Docker, GitHub Actions

COMPLIANCE & OPS

HIPAA, GDPR, PHI data boundaries, audit logging & access controls, Prometheus & Grafana, ELK, chaos testing, HA design

OPERATOR · HOW I WORK

Krishan Saini, side profile

HUMAN_IN_THE_LOOP · the only node that writes the code.

RUNBOOK · WORKING WITH ME

When you hire me, I write the code.

On build engagements there are no hand-offs and no account manager. The person you talk to on the first call is the person who designs the system, writes the code, and carries the pager.

When leadership is the job.

Fractional CTO work means owning the decisions and their records: written strategy, architecture calls with reasons attached, hiring standards, and a codebase your next engineers can walk into. That includes the board and the money side: trade-offs in plain language, build-vs-buy calls with the costs written down, technical risk stated before it gets expensive, and your side of the table in diligence. I stay hands-on enough to be unbluffable. And the job includes making itself smaller: hiring engineers against written standards until the fractional role can hand itself over.

SAMPLE OF THE WRITING · what I’d want a CTO to take from it →

Your engineers get stronger, not sidelined.

On embedded work I review the team’s PRs, write design docs the team can push back on, and pair on the gnarly parts. At VMware, playbooks I wrote were adopted by 100+ engineers because they owned them; that’s the model.

End to end means end to end.

Retrieval to Kubernetes to frontend to analytics: one accountable person, no seams for problems to hide in.

Production is the bar.

Everything ships with caching, autoscaling, monitoring, audit logs, and a rollback path. I spent a decade running systems at 3 a.m., and I don’t build ones that need it.

Constrained models, verifiable output.

A legal model that can’t cite its source shouldn’t answer. A clinical summary a doctor hasn’t signed shouldn’t ship. I design the constraints before the capabilities.

Your repo, your standards.

On client work I ship through your review process, in your CI, matching your conventions, tested and behind canaries. The code stays maintainable by your team after I’m gone.

Inside other people’s systems.

At Pixis I broke a monolith into microservices while product teams kept shipping daily. Brownfield is home turf; most of a decade was spent in codebases I didn’t start.

Built so you never need me.

Client work lives in your repos and your cloud, kept strictly separate from my own products. Decisions and runbooks are written down as I go, so any senior engineer can take over from the docs, not from my memory. The paper terms live in the ledger below.

DOCTRINE · cache before you buy a bigger model · rerank beats raw similarity · the guard ships before the feature · a product isn’t finished when the model works

CONTACT · OPEN A CHANNEL

IST · MY CLOCK

US EAST

EU CENTRAL

Remote, India. Working hours held for US and EU overlap: scheduled, with same-day replies. For production systems I run, I’m pageable; for everything else, those shared hours are the contract.

One engineer. Direct line.

I take on direct client builds, embedded work in existing teams, fractional technical leadership, and principal-level roles. Send a short note about what you’re building, or where your technical strategy is stuck. No spec needed; plain English is enough, and turning it into a system design and a sprint scope is part of the first call. You’ll get a reply from the person who would do the work, and if I’m not the right fit, I’ll tell you who is.

SHAPE01direct client build · embedded in your team · fractional CTO or technical leadership · full-time principal role (employed or contract)
FIRST STEP02short call, then by shape · build: a fixed-scope 2–4 week sprint ending in a clickable build of your core flow at a private URL you can open and share · embedded: a scoped, real ticket through your CI and review, first PR in week one · leadership: a written review of your architecture, roadmap, and process in the first two weeks
BILLING03build work: milestones you can stop at, and every sprint ends with a written scope and cost for the next, so you always know the price of continuing · fractional leadership: a standing cadence we set together, revisited quarterly
RATES04sent in the first reply, with call slots and my current engagement load; same day, no discovery-call theatre
FOCUS05client hours come first, and no engagement is ever dropped mid-milestone. My own products run themselves, with guards (SYS.03, the outbound engine above) · if a full-time role ever starts, in-flight milestones finish first and hand-off runs per AFTER LAUNCH (row 10)
CONTINUITY06the one-engineer risk, named and mitigated structurally: with your repos, milestones, and living runbooks (rows 03, 08, 10), nothing is ever hostage; any senior engineer continues from the docs
HORIZON07fractional is built for years of ownership: milestones de-risk the start; they don’t cap the relationship · leadership transitions are agreed in writing, never abrupt
IP & PAPER08work-for-hire, full assignment · your NDA, your MSA (your standard services contract), your DPA (data-processing agreement) where personal data is in scope · your repos & cloud from day one
VISIBILITY09a working build you can click at every check-in (weekly by default, or the cadence you set), not status reports
AFTER LAUNCH10hand-off with the written runbooks and decision records, or keep me on to operate it (see FOCUS)
SECURITY REVIEW11SIG/CAIQ-style questionnaires (the standard vendor security checklists) answered, not dodged · your SSO, your access policies, least privilege from day one · your device or VDI if policy requires it · your data stays in your tenancy · background verification welcome · solo means solo: no subcontractors; no one but me touches your code, credentials, or data
VENDOR ONBOARDING12insurance & onboarding requirements: send yours with your note; answered line by line, and required coverage arranged to your limits before work starts
AI TOOLING13on your work, only the model endpoints and assistants your policy approves; your code and data never enter tools you haven’t cleared
LEADERSHIP CADENCE14written strategy and decision records as standing artefacts · plain-language trade-off memos for the board · hiring standards and scorecards, all yours to keep
CONTRACTING15via Thinkscoop Technologies (India): my own contracting vehicle and the entity on my resume, not an agency · registration details, tax forms (e.g. W-8BEN-E), and USD/EUR invoicing handled with your vendor onboarding · or via the Toptal network ↗ (verified expert, member since 2021)

Plain-English version: you own everything from day one, work is billed in short steps you can stop at, and the paperwork is yours; I sign your standard contracts.

Measured live in your browser, nothing transmitted; the same habit I bring to everything I ship.