D.S. 第零話 — PROLOGUE
AI ENGINEER  /  INDIA  /  2027

Divyansh Shukla

I build the part of an AI system that says no.

Agents, tool design, durable backends, evaluation. Mostly where agents are allowed near money, which means carefully.

17repositories
7PRs to other people's code
4merged
0double refunds under test
第一話  EPISODE 1

Origin

In which the rule is established.

01 / RECORD

Subject fileKept short on purpose.

NAME
Divyansh Shukla
OCCUPATION
AI / LLM engineer. Research engineer.
LOCATION
India. Remote, Bangalore, or anywhere.
STATUS
Final year. Full-time from 2027. Internships and remote contracts before that.
TRUSTS
Provider state. Test logs. Receipts.
DOES NOT TRUST
Timeouts. Screenshots. His own first numbers.

Most AI demos end at "the model said so." I work on what comes after: the retry after a timeout, the reminder sent to a customer who already paid, the post that quotes a number the code doesn't support.

So I build the dull, durable layer around the clever part. Then I try to break it before someone else does.

THE RULE

The model says what must be true. Live state says what is true. Code decides.

第二話  EPISODE 2

Records

Things that were built, and what they were built to refuse.

02 / WORK

RecordsFrom github.com/devdiv07. Ordered by consequence.

01LIVE

PREFLIGHT

Stops recovery agents from asking customers who already paid to pay again.

An AI risk gate for payment-recovery agents, built for Razorpay AI Buildathon Track 02. The model identifies what must still be true. A live read of provider state checks it. Deterministic policy returns ALLOW, BLOCK or ESCALATE. ※ The first result was +0.556. A self-audit found a segmentation bug in the baseline. The corrected figure is +0.302 recall (0.889 vs 0.587), reported as a diagnostic, not a headline. PYTHON · RAZORPAY TEST MODE · POLICY ENGINE

02LIVE

Financial Operation Core

A timeout is not a failed refund.

Durable execution for money moves that agents start. A naive retry produces two refunds; this produces one. That holds through lost responses, crashed workers and 20 concurrent callers, measured in Razorpay Test Mode. The Chaos Lab lets you try to make it refund twice. PYTHON · IDEMPOTENCY · CHAOS TESTING

03OPEN SOURCE

Shipyap

You ship. It yaps. With receipts.

Turns GitHub work into LinkedIn and X posts with repo citations. A draft that quotes a number the repo can't support gets blocked. One human review, then Buffer publishes. About $0.15 a draft. PYTHON · CLAUDE API · BUFFER

04ML

Multimodal Message Router

Reads, listens, looks. Then routes.

BLIP captioning, RapidOCR and Whisper behind strict provenance validation. 0.966 action macro-F1 under leakage-controlled evaluation, with a four-mode ablation harness. BLIP · WHISPER · OCR · ABLATIONS

05VISION

ClaimTrace

No free text allowed.

Multimodal claim verification. GPT-4o returns only Pydantic-validated structured output, wrapped in deterministic rule checks. GPT-4O · PYDANTIC

06RESEARCH

When is a screenshot evidence?

A question, asked properly.

Research on capture-time provenance: when is a screenshot from a computer-use agent actually enough evidence of what the agent did? AGENT EVALUATION · PROVENANCE · REPRODUCIBILITY

07LIVE

Attesta

Compliance, written down where it can't be edited.

Checks financial content for SEBI risk with Llama 3.3-70B. Attestations go on Ethereum Sepolia, full reports on IPFS, and compliance badges are minted as NFTs. NEXT.JS · SOLIDITY · SUPABASE · IPFS

08AUTOMATION

n8n-based CRM

Nine workflows. No one has to remember to follow up.

Lead intake, qualification, booking, follow-up, reactivation and reporting, built in n8n with OpenAI, Google Sheets and email. N8N · OPENAI · WORKFLOWS

09IN PROGRESS

MIRROR

No buy or sell calls. No promises.

A source-linked research assistant for an India and US stock watchlist. Each morning it reports what material events happened and why a watched stock moved, with dated evidence and what is still unknown. The first slice is tested on synthetic fixtures. PYTHON · RESEARCH TOOLING

第三話  EPISODE 3

Other people's worlds

Pull requests to code he doesn't own.

03 / OPEN SOURCE

Pull requestsSeven opened. Four merged. Two still waiting.

第四話  EPISODE 4

Method

What he actually knows how to do.

04 / METHOD

TechniquesEach one was used in a record above.

AGENTS
MCP servers
Tool contracts
Agent-to-agent delegation
Structured outputs
DURABILITY
Idempotency
Retries after lost responses
Crash recovery
Chaos tests
EVALUATION
Ablations
Leakage control
Confidence intervals
Auditing his own numbers
MULTIMODAL
Vision LLMs
Whisper
BLIP
OCR
TRUST
TPM 2.0
Intel TDX
Capture-time provenance
On-chain attestation
STACK
Python · TypeScript · Go
Next.js · Supabase
n8n · Solidity

No skill bars. Nobody is 87% at Python.

第五話  EPISODE 5

Scale

The universe is very large and almost entirely untested.

最終話  FINAL EPISODE

Your move.

Open to AI/LLM engineer and research engineer roles in agents, evals and safety. Full-time from 2027, internships, or remote contract.