Sabyasachi MishraAI Product Manager

I build the system around the model.

I turn ambiguous AI opportunities into production systems with explicit evaluations, handoffs, and guardrails.

I also operate independent products and infrastructure, so the learning survives contact with reality.

Inspect the work Read the thinking

Sabyasachi Mishra outdoors in the mountains

Current fieldAI × product × autonomy

Voice AI40K

leads handled monthly

Vision systems48×

faster audit cycle

Live compliance22

operating cities

Enterprise NLP₹40K Cr

portfolio monitored

Track record

Stanza Living / Power Finance Corporation / Brahmagupta Edu / IIM Indore

01Point of view

The model is a component.The product is the operating system around it.

Reliability lives in the boundaries: what good means, when the system should act, when a person should step in, and how the product learns after launch.

An editorial system map connecting evaluation, human handoffs, guardrails, and production learning.

Define good before writing code.

An AI feature without an evaluation set is a demo with a roadmap. I establish the measurable boundary first, so the team knows what should ship and what should not.

Design the handoff, not just the happy path.

Confidence routing is a product decision. Models handle what they can; people handle what they cannot. The quality of that boundary often decides whether the system earns trust.

Guardrails belong in the architecture.

PII masking, access control, tone constraints, and adversarial thinking shape the system from day one. Trust is not a QA ticket at the end of a sprint.

Treat drift as a product bug.

Production models degrade quietly. Monitoring, feedback loops, and retraining triggers need owners and thresholds just like any other critical product behaviour.

02Selected systems

Proof through the work.

Three production systems shaped by uncertainty in conversation, ambiguity in visual evidence, and trust in regulated decisions.

01Voice AI · Lead conversion

Riya voice platform

A voice agent designed to know when the conversation should no longer belong to the model.

Riya handles a high-volume lead journey in the language people actually use: fluid Hindi-English conversation, with explicit routes back to a human when confidence or sentiment drops.

40Kleads handled each month

60 / 40Hinglish code-switching pattern

HITLconfidence-based routing

Inspect the system

System logic

01Leadintent + context
02DialogueHinglish voice loop
03Confidence gatesignal + sentiment
04Next actionconfirm or nudge

Low confidence → human sales handoff

LLMVoiceHuman in the loopDrift

02Computer vision · Operations

Vision-based compliance

Turning a 12-hour inspection lag into a near-real-time operating signal.

A VLM-based audit loop evaluates housekeeping evidence, routes uncertain cases to review, and gives operators a daily view of compliance across a distributed network.

48×faster audit cycle

+35%increase in coverage

400+stakeholders on live data

Inspect the system

System logic

01Evidenceroom images
02VLM auditchecklist evaluation
0365% gateconfidence threshold
04Ops viewstatus + exception

Below threshold → human quality review

VLMComputer visionCPU inferenceMonitoring

03Enterprise NLP · Credit risk

Risk intelligence platform

Making early-risk signals legible inside a regulated, document-heavy workflow.

An NLP system surfaces risk events from unstructured financial documents, then delivers them through role-aware interfaces designed for the people accountable for the portfolio.

₹40K Crportfolio monitored

85%document parsing automated

~20%risk exposure flagged early

Inspect the system

System logic

01DocumentsPDFs + audit notes
02NER layerrisk entities
03Policy layerPII + access
04Reviewofficer workflow

Ambiguous signal → analyst verification

NLPNERPII maskingRole-based access

03Operating in public

Public products. Visible decisions.

Three live services are secondary evidence of end-to-end product ownership: choosing the boundary, shipping the interface, and living with the operating consequences.

P/01 Live

Visual feedback for coding agents

Shiproom

“The button feels broken” is not a bug report. Someone still has to turn it back into a route, an element, and a browser — usually afterwards, without the screen in front of them.

Product decision

A PM owns the quality of the evidence handed from a user to a builder, not only the interface that collects it.

Open shiproom.live

The Shiproom reviewer demonstration: a sample checkout page with two numbered reviewer marks on the “Pay now” button and the estimated tax row, beside the context Shiproom attaches automatically — route, page title, selected element, accessible name, viewport, browser, screenshot, mark types, and a failed POST request that returned 500. — The public reviewer demonstration on shiproom.live. Two marks on an invented checkout, and the context that leaves with them.

P/02 Live

Local-first teleprompter

teleprompter.wtf

The reliable thing must not depend on the clever thing.

Product decision

Reliability starts with a useful baseline that survives when optional intelligence, permissions, or the network do not.

Open teleprompter.wtf

The teleprompter.wtf script editor holding a short neutral rehearsal script, showing a live count of 98 words and an estimated 45 seconds at 130 words per minute, a “Saved only on this device” indicator, a speaking-pace control, and the Start teleprompter action. — The editor, with local word and time estimates.

P/03 Live

Precise image workspace

compressimage.fun

An exact size limit is a search problem, not a quality slider.

Product decision

Precise promises require an operating method that proves when a target was met and says clearly when it was not.

Open compressimage.fun

The compressimage.fun result workspace after an exact-size run on a synthetic test image: 69.9 KB reduced to 48.7 KB under a 50 KB cap with the original 1600 by 1067 dimensions kept, an original-versus-processed preview toggle, follow-on resize, crop and convert actions, a Delete now control, and a note that files delete automatically. — One exact-size run on a synthetic test image: 69.9 KB down to 48.7 KB under a 50 KB cap, dimensions intact.

04Independently operated infrastructure

I operate the edge cases myself.

Operating the runtime, memory, deployment, and observability is how I learn where autonomous systems fail outside the demo.

04Personal infrastructure

OpenClaw gateway

An always-on agent boundary designed to be easier to reach than to over-permission.

24/7always-on gateway1phone number allowlisted

05Personal infrastructure

Private agent memory

An agent remembers more usefully when its memory is treated as private infrastructure, not a rented feature.

768local embedding dimensions~2GBquantized extraction model

06Personal infrastructure

Content operating system

A content pipeline where deterministic code owns state, models own judgment, and people retain approval.

2parallel pipelines8+staged workflow states

What I run myself

I own the operating consequences across this stack. That makes failure modes visible before they become abstract product advice.

Servers and hosting
Containerised deployment and releases
Domains and production routing
Agent runtime
Private memory and retrieval
Model selection and routing
Permissions and security boundaries
Monitoring, analytics, and failure handling

05Smaller experiments

Questions with working interfaces.

Each one existed to answer a focused product or systems question.

L/01 Shipped

Natural-language data access

An NL2SQL copilot that turns a business question into a constrained path through internal operational data.

The interface is conversational; the product work is permissioning, schema context, and verification.

NL2SQLInternal tools

L/02 Running

Hook-pattern learning loop

A feedback layer for the content engines that records performance at 2h, 24h, and 7d, then turns useful themes and structures into inputs for the next draft.

A learning loop should record uncertainty honestly—unavailable is not zero, and fabricated data teaches the wrong lesson.

OpenClawObsidianFeedback loop

L/03 Shipped

Small models for high-frequency jobs

A fully local memory loop using a 3B extraction model, local embeddings, and Qdrant to keep agent recall private, cheap, and continuously available.

A frontier model belongs in the conversation; a smaller model can quietly own the narrow, always-on background task.

OllamaQdrantmem0

L/04 Testing

AI video pipeline

A production pipeline exploring where generative tooling can compress repetitive video operations without hiding quality decisions.

The unit of automation is rarely the whole workflow; it is the slowest repeatable decision inside it.

MultimodalWorkflow

06Thinking

Ideas sharpened by building.

Notes on the product decisions that sit between a capable model and a dependable system.

Essay2026 · 5 min

AI-native transformation, not another chatbot

A contrarian case for rebuilding operations, decisions, customer experience, and internal workflows around AI—not attaching another bot to the company that already exists.

Essay2026 · 5 min

The pipeline is the product

What building an autonomous content operating system taught me about where automation should stop, and why the model is only one component of a reliable workflow.

Essay2026 · 4 min

Memory makes the agent

Why I self-host an agent’s memory on a small local model and a private vector store—and why memory is a product sovereignty question, not an API checkbox.

NowUpdated August 2026

Operating the gap between autonomous action and accountable products.

Current attention: keeping public products and private agent infrastructure running, sharpening evaluation systems and human handoffs, and making autonomy safe enough to be useful.

See what I'm exploring

07Connect

Bring me the consequential product problem.

If you are turning an AI prototype into an accountable product, working through an evaluation or handoff, or need a product leader who can operate across strategy and systems, tell me what is at stake.

Email[email protected]NetworkLinkedIn CodeGitHub Short notesX / Twitter