Flag · tracker vs reality
Project rated fully healthy in the tracker. The chat shows the client actively reassessing the engagement.
surfaced by cross source audit · routed to a human
Amaan Abdul Kader · Space
This is Space.
One platform runs the company's mornings: a nightly pipeline that reads every client and vendor conversation, a second model that audits the first, evals that decide which model earns each job, and the tools that grew around it. This page is the whole story, stage by stage.
System nightly
87 projects
Six stages
Observatory time ··:··:·· IST
03:14 · nightly sweep started · 87 projects queuedrepresentative recreation
The engine
Six stages between chaos and a clear morning.
The flagship build: a nightly pipeline that turns tens of thousands of unstructured WhatsApp messages into one send ready action brief per project. Facts are extracted once, at ingestion. The morning dashboard reads plain database rows, opens instantly, and never calls a model just to load a page.
01 · Ingest
Every conversation flows in overnight.
A resilient ingestion layer with session isolation, dedup reconciliation, and watchdog recovery. It has survived every flaky reconnect the real world has thrown at it.
02 · Compute
Deterministic health scoring. Zero model calls.
Five weighted components score every project in the portfolio before any AI wakes up. Nothing at this layer can hallucinate, because nothing at this layer is generative.
03 · Sweep
Small models extract structured facts.
Schema locked output with code level guardrails catching what prompts cannot. The cheapest model that passes the benchmark earns the job, and nothing more expensive.
04 · Brief
One brief per project, ready before breakfast.
Health score, ranked actions, send ready client drafts, team directives with names and deadlines, vendor commercials. Every fact carries a link to the exact message it came from.
05 · Audit
A second machine rereads the first one's work.
A separate model sweeps the night's output against every source and raises what a human reviewer would miss. 147 cross source flags on its first full company run.
06 · Weekly
The long exposure.
Weekly digests, portfolio patterns, team level trends. The constellation view of the whole book, assembled from a week of nights.
87
projects, nightly
147
flags, first audit run
$14
a night, full sweep
0
model calls to load a page
The evidence
Software that notices.
Dashboards report what people typed in. This system reads what actually happened, compares the two, and raises its hand. The kind of thing it catches:
Representative recreations · details redacted · no client data appears anywhere on this page
Flag · finance
Payment from received weeks ago, never logged as collected. The books said zero.
chat evidence linked · corrected same day
Flag · stale value
Contract value unchanged in the CRM after a scope reduction agreed in chat. The system quoted the message.
receipt attached · value corrected
Flag · silence
Client silent 68 days with 11 decisions pending. A recovery play was prescribed in the morning brief.
silence computed from raw timestamps, not memory
And what it refuses to do.
A project with zero client messages is forced to unknown sentiment. A hard directive forbids inference and the schema enforces it. The system is not allowed to guess, because one early brief guessed, novelized a client's mood from internal chatter, and taught me the whole class of bug.
And what it was not allowed to ship.
A downstream action auditor measured at 55 to 60 percent precision. It was paused, not shipped. A number you can trust is worth more than a feature you can list.
The proving ground
No model gets a job here without winning it.
The eval harness: 22 scenarios built from real operations, including adversarial traps, scoring JSON reliability, schema pass, grounding, and hallucination count. Judged blind by a model from a second vendor, so nobody grades their own homework. More than ten models have competed.
Thinking modes
Thinking modes hurt structured output. Every extended thinking variant tested raised parse errors on extraction work.
Model routing
Cheap models, fenced by guardrails written in code, matched frontier models on extraction. Each pipeline stage now routes to whichever model earned it on the bench.
Champion rule
Champion vs challenger before any swap. The latest pipeline beat the incumbent 6 to 0 in blind judging and cut hallucinated figures from 12 to 2.
Unit economics
Routing, prompt caching, and batch pricing brought the full nightly sweep to roughly $14. The frontier tier is reserved for the one layer that earns it.
Locked before anyone was invited.
Before a single colleague saw the dashboard, I audited the database it runs on. Most tables were readable by any authenticated login. Two were open to the public internet with nothing but the anon key.
Everything went into a tiered row level security model. Then it was verified the only way that counts: by trying to break in.
The blocked login below was the security model passing a real world test.
Security verification · before multi user launch
» simulate member token, cross portfolio read ··· 0 rows
» platform security advisor scan ··· clean
» non allowlisted executive login ··· blocked
» allowlist 17 accounts ··· 14 active within 2 days
» platform security advisor scan ··· clean
» non allowlisted executive login ··· blocked
» allowlist 17 accounts ··· 14 active within 2 days
The wider constellation
Tools that gave people their hours back.
Around the engine, a suite of live, gated, self serve tools, each one replacing a manual process that used to cost a person an hour. All designed, built, shipped, and operated solo, alongside the day job.
Venue Calculator
A real inquiry goes in. Every venue comes back fit ranked with holding cost pricing, comparison, and print.
634 venues indexed
Moodboard Engine
A couple's taste becomes a four event moodboard book with per event palettes and a print ready export.
128 image reference library
Cost Sheet Studio
Builder on the left, client ready proposal on the right. Preserves real pricing strings instead of inventing precision.
0 fabricated figures
Welcome Guides
Couple facing guides with branded short links, usage telemetry, and an audit logged admin surface.
10 versions shipped
Image Library
Upload, tag, search. Images resized in the browser before a single byte leaves the client.
engineered inside a 12 function platform cap
Local Signal Lab
A 4 billion parameter model on a 6GB laptop GPU, benchmarked on a 14 trap suite, so sensitive chats never have to leave the building. A delta contract cut model calls on a busy day from 49 to 8.
93% on the trap suite
The operator
The systems come from the floor.
I lead four planning teams and eight direct reports at a destination wedding company, across more than 900 celebrations. I have run escalations after midnight, recovered clients mid crisis, and negotiated across a 500 vendor network. Then I did the unusual part: I encoded that judgment, 36 standard operating procedures and 52 documented warning signals extracted from my own practice, and made them the rubric a machine critiques every brief against. The AI is good because the judgment inside it was earned first.
Age 19.
Led a Guinness World Record build: the longest cake ever made, 5,300 meters, 100 baker teams, an 8 month project.
2024.
Best Wedding Planner, national user choice award from India's largest wedding platform.
11 months.
From relationship manager to Team Lead, now running the destination planning function's portfolios and people.
9 years.
In events and operations, started at 16, every year of it client facing and unforgiving of excuses.
Wedding planner by title. Systems builder by evidence.
- No claim without a receipt.
- Guardrails live in code, not prompts.
- Honest uncertainty beats confident fiction.
- Nothing ships without beating the champion.
The observatory is open.
If you want great software built, whatever shape it takes, I would like to talk.
Jaipur, India · open to Delhi NCR, Bengaluru, Mumbai, relocation, remote · available within 30 days