Forward deployed engineering

Most AI systems that look good in a demo fail quietly once real users and real data arrive. Forward deployed engineering is the work of building the system inside a team's actual data and workflow, then measuring whether it holds. I have done that at Lenskart and at C3alabs. The full records below give the users, the problem, what I built, one decision and its tradeoff, and the measured result.

  • 54% → 89%

    LLM-judge agreement with human labels, two calibration batches

  • 35 to 42%

    cost cut per request on live runs, AI workspace for wealth managers

  • 91.8/100

    Soundstage reel score from a separate AI reviewer

Lenskart, 2026.

Built end to end.

Competitor creative intelligence and its eval layer

Context

Lenskart is an eyewear company that runs large ad campaigns. Its leadership wanted a reliable read on what competitor brands were running.

Users

Lenskart leadership, through an internal analyst.

Problem

The earlier pipeline was measured making up facts in 27% of sampled rows. Leadership could not act on numbers it could not trust.

Built

A pipeline that collects competitor posts and ads from Instagram, TikTok, Meta Ads and YouTube, tags each one on 24 structured dimensions with Gemini, and shows the results in a dashboard with a chat assistant. It runs on a schedule across a set of competitor brands. A deterministic guardrail checks every number the assistant writes against the source data.

Decision

Tagging calls ran one at a time at first and hit the model provider's free-tier limit of 20 requests a day. I moved them to one combined batch job per cycle across all brands. The tradeoff is freshness: new ads get tagged at the next cycle, not instantly, so that the job stays inside the quota.

A judge metric measures how often an automated grader agrees with human labels. It shows how far to trust the grader. It does not show how often the agent is right.

Measured

The LLM judge agreed with human labels on 54.0% of 50 held-out real user turns in an early version, and on 88.9% of 27 unseen turns in a later batch, after rounds of calibration. The two figures come from different batches.

Lenskart, 2026.

Team project; I built the parts below.

Hiring intelligence platform: search, recovery and identity

Context

Lenskart's recruiters hire for technology and leadership roles and needed a faster way to find and rank candidates from the open web.

Users

Lenskart's recruiting team.

Problem

Candidate search ran long jobs across several AI providers. A crashed worker or a failing provider lost the search, and loose matching risked merging two different people into one profile.

Built

Candidate search across 4 LLM providers, each grounded in its own live web search, with a switch to another model when one fails. A Postgres lease and heartbeat system hands a crashed worker's search to another worker, and the search runs as a scheduled background job. Ranking uses 4 explicit keys (score, evidence coverage, confidence tier, name) instead of one opaque AI score.

Decision

We stopped using name and company matching as proof that two profiles are the same person, because it could merge two different people. The accepted cost is that one person found under two profile URLs can appear twice. Unclear matches are recorded for later human review.

Measured

No figures are published for this layer yet.

Lenskart, 2026.

A short entry: this is an internal product and its details stay private.

Leadership intelligence brief: ingestion and evidence layer

Users

Lenskart leadership.

Built

The ingestion and evidence layer: source connectors that write into one Postgres system of record, read by the curation layer through an MCP (Model Context Protocol) server.

Grounding

Rules that decide when a claim has enough independent sources to be used, and that reject any name the source text does not contain.

C3alabs, Oct 2025 to Oct 2026, part-time.

Team codebase; I built the subsystems below.

Cost and reliability in an AI workspace for wealth managers

Context

C3alabs builds an AI workspace for wealth-management and financial-advisory firms. It turns client meetings, calls and documents into notes, follow-ups and CRM updates, and runs workflows across connected tools.

Users

The product is built for advisors and operations teams at those firms.

Problem

Each request carried large prompts, so cost per request was high, and some failures in the agent workflow engine did not show up as errors.

Built

I cut the classifier prompt from 19,141 to 5,406 characters and the planner prompt from 22,779 to 9,482 characters, which saves about 3,400 and 3,300 tokens per request. I found and fixed 21 defects across the 3-repo workflow engine (13 blocker or high, 8 medium or low).

Decision

The prompt cache was being invalidated on nearly every request, because a memory block that changes with each query sat above the cache boundary. I moved it below the boundary. The memory block is no longer cached, but the large stable part of the prompt now is, on every request.

Measured

Live runs reported 23.6K tokens cached per request and a 35 to 42% cost saving; other live logs showed savings of 25 to 35%. In a local verified run, a 3-step task needed 1 tool-router handshake instead of 3. A tool-permission bug that made the agent loop for 15 turns was fixed, and the agent no longer blocks any tool calls in that run.

Products built end to end

  • Soundstage

    Built solo for the Gnani Great Indian AI Internship Challenge. It is for Indian D2C brands and small agencies: a brand records a voice note in Hindi, Hinglish or English, and gets an on-brand 9:16 reel that a separate AI reviewer scores before anyone sees it. A 14-second Hinglish note became a reel scored 91.8/100 in 3 minutes 36 seconds. Across 59 measured runs, the median run took 6.1 minutes, and the median model cost of the agent that plans and writes the reel was $0.46. It runs on demo brands.

    Soundstage case study

  • Outfox

    Built for the SerpApi India Hackathon 2026, for marketing teams, starting with Indian D2C beauty brands. It watches competitor ads, videos, search and news, tags each finding by hook and audience stage, and links every claim to the public page it came from. An Ask agent answers only from stored evidence. It is a public demo with no usage figures claimed.

    Outfox case study

The first 30 days on a customer deployment

  1. Week 1

    sit with the people who will use the system. Collect 30 to 50 real tasks and the failures they care about. Write down what "working" means as numbers.

  2. Week 2

    turn those tasks into an eval set and record a baseline. Ship the thinnest end-to-end path on their real data.

  3. Week 3

    improve against the evals. Add guardrails where a mistake is expensive: grounding checks, approvals, cost limits. Trace every run.

  4. Week 4

    hand over a regression gate in CI, a runbook, and a written record of what changed and what still fails.

Open to applied AI and forward-deployed engineering roles. Contact

“A man who is master of patience is master of everything else.”

~ George Savile

© 2026 Mohit Goyal