Agents that run unattended
Scheduled systems that do the work and report back. Three of mine run every day without being started by hand.
SUNDAY, Chief of Staff, Stock News Monitor
Mark One
I build AI systems that keep running after I close the laptop.
Mark One is a one-person AI studio in New Delhi, run by Ankit Keshari.
What I build
Three things, and they are outcomes rather than technologies. Each one has systems behind it that already run.
Document and research pipelines
Scanned or scattered material in, structured and cited output out. Every score quotes the source line it came from.
The method behind the EY government platform
Local-first and private tools
Runs on the machine. Nothing leaves it. No subscription, no cloud round trip, no renting you your own voice.
Aadesh, JARVIS
Products
Three builds shipping in the next one to two months. A build number is earned by shipping to someone who is not me.
Mark I
Scout
Finds the work, ranks it, applies on your word.
Scrapes openings daily, ranks them against your actual resume, builds ATS-safe PDFs and emails you the top five. Reply with one word and it applies. It never contacts the same company twice.
Mark II
Aadesh
Hold a key, speak, it types. Nothing leaves the laptop.
Silero VAD gates the mic, faster-whisper transcribes, and a cleanup model polishes the text before it lands at your cursor in whatever app is focused. Local by default, or your own API key if you want the cloud.
About 1.5 seconds end to end for short utterances on an i5-1235U with no GPU. Zero network calls in local mode.
buildingNotify meMark III
Cockpit
The state of your day, at a glance.
An always-on desktop window that shows what is happening and never interrupts you about it. Built on Tauri, so it survives a closed terminal and can write to the filesystem.
Systems in production
Live means it runs and is reachable. Three of these also run unattended, on a schedule, without being started by hand. Repositories are linked where they are public.
live5
SUNDAY
Job-hunt agent. Scrapes openings daily, ranks them against a real resume, builds ATS-safe PDFs, emails the top five, and applies on a one-word reply.
Live daily at 10:00 IST. Has never applied to the same company twice.
RepositoryStock News Monitor
Scores every headline and company pair from Indian financial feeds, then grades itself against the next trading day.
Reports excess return against NIFTY, split by after-hours and in-session, novel and descriptive.
RepositoryChief of Staff
Telegram ops bot on Cloudflare Workers. Routines, SM-2 spaced repetition, and an 8-8-8 hour budget.
Hard cap of two unprompted messages a day.
RepositorySunday, Let's Get a Job
Placement-strategy app. Company to rounds to questions, with readiness scoring.
Single user, PIN-gated, and zero rupees a month in infrastructure.
Portfolio
One shared WebGL canvas behind a single dark page, degrading to clean static HTML without it.
Live. First load around 168 kB before three.js.
Repository
building2
JARVIS
Local voice layer. Wake word, local transcription, and confirm-before-act on Windows, Spotify and WhatsApp.
Wake word runs at roughly 0.02% of a core. Nothing leaves the laptop before the wake word fires.
Agent Router
Three-lane delegation. One model decides, cheaper models type, and everything comes back for proofread.
Both lanes verified end to end. The same latent bug was found in both.
private1
Multi-Agent Board
Tauri desktop board for watching agent sessions, with hand-built SVG instrumentation.
Built. Not published.
Client and institutional work
Work an external party commissioned. As of today that is exactly three, and nothing else on this site is described as client work.
EY
AI Intern
Jun to Jul 2026
LLM document-intelligence platform for a government client. OCR over scanned official journals by page rendering plus vision-model transcription, cached so nothing is read twice. Rubric-driven scoring where every score quotes the exact source line it came from.
Evaluated against 136 expert-rated products, with exact-match and within-one accuracy tracked per parameter to choose the model.
FastAPI, PostgreSQL and pgvector, LlamaIndex, Gemini Vision, Next.js
Described at the level of method. No client data, no rubric contents, no scores.
DRDO, Scientific Analysis Group
Research Intern, Adversarial ML
May to Jul 2025
Black-box adversarial attacks on face recognition, to measure how the models break. SimBA and NES pipelines, then transfer-based PGD with norm constraints.
Across 8 or more models and 9,000 or more samples, with transferability measured across architectures using DeepFace embeddings.
Python, PyTorch, DeepFace, NumPy
MUN Prep Platform
Commissioned build
2026
Evidence-backed Model UN preparation for a class-10 delegate. Committee, country and agenda in, one humanized Word document out: country dossier, position paper, GSL speech, caucus speeches, points of information, rebuttals and draft clauses.
Citations tiered, separating official primary sources from reputable secondary ones.
Building next
CodeLens
Write DSA code and watch it execute. Memory, stack, heap and recursion, live as you type.
A trace format plus a language-agnostic visualisation engine. Languages are plugins.
ECG Interpreter
Educational ECG interpretation on PTB-XL. A classifier decides, an LLM only explains.
The separation is the safety design. The model that writes the words never sees the waveform.
AttendAI
Face-recognition attendance, recognition spike first, SaaS scaffolding deferred behind explicit triggers.
The whole plan hinges on one unvalidated assumption, and says so.
About
I am a B.Tech student at NIT Delhi, class of 2027, with a mechanical engineering background that shows up in how these systems are built: tolerances, control loops, failure modes, and a bias toward degrading rather than breaking.
Two internships behind me. Adversarial machine learning research at DRDO's Scientific Analysis Group in 2025, then an LLM document-intelligence platform for a government client at EY in 2026.
Two rules run through everything here. It has to ship, and the numbers have to be real.
What do you do every week that feels the same?
You do not need to know what can be automated. Describe the part of your week that repeats, or the thing only one person knows how to do, and I will tell you straight whether a system helps. Sometimes the honest answer is that it does not, and I will say that too.
Or email
ankitkeshari2006ai@gmail.com