Kripa Shankar Mishra

Top of page 0%

Systems · evidence · boundaries

Building what excites me.

I design and build AI products where trust, retrieval, privacy, and durable state need to survive contact with real systems.

India · engineering portfolio · public-safe by default Read the work index →
Engineering notes and system diagrams arranged on a desk The record of the work starts before the model: notes, decisions, and evidence.

01 / Selected work

Small surfaces. Serious boundaries.

Architectural threshold with a single red handoff marker
Editorial illustration — a threshold with one red handoff marker, used for the Career Copilot case study, which documents a single-owner agent with a bounded /save flow.
Featured case study · in-development Career Copilot A single-owner Telegram career agent: submit a job URL and get a private, PII-redacted Markdown fit report saved to Google Sheets. The public README documents a bounded /save flow, server-created request context, and owner-scoped persistence. Multi-user and browser-assisted features are explicitly out of scope. Read the case study →
02

mastra-pii

PII redaction for Mastra agents as an adapter, not a detector: one interface over a Presidio deployment — the package owns integration, the deployment owns intelligence — with a deterministic Indian-recognizer layer and fail-closed defaults. 03

DocNeedle

Local-first SQLite FTS5 CLI for deterministic doc indexing and retrieval — index, search, read, and stale-term check across markdown corpora, zero LLM/API in the retrieval path. 04

Sales Intelligence Agent

FastAPI and LangGraph sales-intelligence API that answers structured and causal questions over a cleaned SQLAlchemy star schema plus business-note retrieval, with citations, semantic caching, async queueing, and golden-set evals.

Four areas carry most of this site's agent-systems work.

02 / Current focus

Confirmed in the case studies: bounded context, honest evaluation, privacy-aware workflows, and fallbacks that work.

Every one of the four areas has at least one project record or ADR behind it on the Work index.

  • Memory and context boundaries
  • Evaluation tied to user value
  • Privacy-aware agent workflows
  • Fallbacks that preserve useful outcomes

Read the work index →

03 / Writing

Writing from real project decisions.

Each essay starts in an existing README or accepted ADR and links back to the project that raised the question — decisions with their evidence, not standalone opinions.

Aug 18, 2026 agent-systems

An eval harness for a user-facing agent

An agent chatbot cannot be unit tested like a normal program. Career Copilot splits evaluation into a deterministic contract lane that blocks merges and a budgeted quality lane that never overrides it.

Aug 12, 2026 pii-redaction

Drop-in PII redaction for Mastra agents

Agent loops touch sensitive text at every boundary — prompts, tool calls, traces, memory. The hard part is not detection; it is integration.

Aug 10, 2026 developer-tools

Separate the product surface from the benchmark harness

DocNeedle's README is persuasive because it treats normal CLI use and repo-specific evaluation as different things instead of collapsing them into one story.

Aug 10, 2026 agent-systems

Trusted ingress before more tools

Career Copilot shows why a single-owner agent benefits from tight trust boundaries before it grows new capabilities.

04 / Resume

A verified experience timeline.

Roles, dates, and evidence-backed highlights only — no unsupported location or work-mode claims.

Open the timeline → Read the portfolio framing