01 / 05   link established

Open to offers · — notice HSR Layout, Bengaluru Onsite · Hybrid · Remote

SEC.01 / IDENT

Sagar Chaudhary

AI engineer for industrial and manufacturing systems.

I take high-frequency sensor data off an engine test bed over raw TCP/IP, put it on a live dashboard in under a second, train LSTM models to call the failure before it happens, and hang a RAG layer on top so a plant engineer can just ask.

Most AI work stops at the notebook.

I start at the sensor.

Raw TCP/IP off the test bed. Sub-second to the dashboard.

LSTM forecasting on the live run — the anomaly is flagged before the part fails.

Then a RAG layer over the test reports, so nobody digs through a PDF again.

Signal to decision. That is the whole loop.

SEC.03 / SIGNAL

Anomaly detection, running here

Trailing-window mean and σ, EWMA-smoothed z-score, threshold with hysteresis — and the same buffer through a 256-point FFT, in orders of shaft speed. The detector shape that runs on the test bed.

VIB-01 · live window 110 · ewma 0.28
NOMINAL
State
0.00
z-score (ewma)
0.000
σ, trailing
3.60 σ
Threshold

Press inject. A bearing-defect signature is added to the trace: an impulsive high-frequency ring that grows and decays. For the first few samples the amplitude still sits inside the envelope — you cannot see it. The z-score already can.

That gap, between when a detector knows and when a human would, is the entire commercial case for predictive maintenance. Turn the sensitivity up and it fires earlier and false-alarms more; turn it down and it goes quiet. Choosing that number is most of the job.

Now switch to spectrum and inject again. The same buffer, through a 256-point FFT, plotted in orders of shaft speed. The fault that was invisible in the time trace stands straight up at 7.2X. Rotating machinery is diagnosed in the frequency domain, not the time domain — this is why.

Synthetic signal, real detector. Plant data stays with the plant.

SEC.04 / ARCHIVE

A sample a second, for hours

What a historian actually stores, and the error bound it promises in exchange. Swinging-door compression, running here.

TE-204 · 30 min run swinging door · tolerance ±6.0 °C · corridor ±3.00 °C
1800
Samples in
0
Points stored
0:1
Compression
0.00 °C
Max error, measured

The pale line is every sample the thermocouple produced. The bright line is what a trend client draws back from the handful of points actually written to disk — the sodium dots. Widen the deadband and watch the ratio climb.

The promise is worth stating precisely, because it is easy to get wrong. The door swings inside a corridor of half-width E from the last stored point — but the line drawn back joins stored samples, each of which may sit E off that corridor, so the reconstruction is bounded by 2E. The slider therefore sets the tolerance you are willing to accept, and the corridor is half of it.

Max error is measured, not asserted: the chart rebuilds the reconstruction and compares it against all 1800 originals. A test over 240 generated runs holds it under the stated tolerance every time.

Decimation would keep every tenth sample and lose the transients. This keeps the corners and throws away the straight runs.

SEC.04 / SYSTEMS

Nine systems, shipped

Six built inside the role at Deevia Software, three independent. Open one for the engineering detail and the exact stack.

BENCH / admission control - Inferno batch 8-or-5 · HIGH 60 · LOW 24
ADMITTING
Gateway
115
req/s in
0
Queue depth
0
429 sent
REQ/S - crank it drag the knob, or arrow keys

This is Inferno's admission control, operable. Crank the load past what the workers clear and queue depth climbs; at the HIGH watermark the gateway sheds with 429 + Retry-After, and it only re-admits below LOW.

The gap between the two watermarks is hysteresis. Without it the gate chatters at the boundary; with it the system settles. Same latch as the anomaly detector above - it is the recurring trick of stable systems.

Poisson arrivals, size-or-timeout batching, dual watermarks - the same mechanics as the production bullets, one screen up.

SEC.05 / CAPABILITY

What is actually in the toolbox

Compiled from the same profile the resumes are generated from. Nothing appears here that is not used in a system above.

SEC.06 / TRAJECTORY

How it got here

Seven-month internship converted to full-time. Sixteen months of production AI experience as of today.

Trend cursor - drag to interrogate any month

Availability

    Resume — six variants

    Same facts, different emphasis. Section 07 picks the right one for a given job description.

    Annunciator - recognition & certification. Flashing tiles are unacknowledged; click to ACK. An acked certificate opens its verification page at the issuer.

    SEC.07 / MATCH

    Paste the job. Get the honest answer.

    Coverage against the real vocabulary, the gaps named out loud, and which of the six resume variants this role should get.

    INPUT / job description Ctrl/⌘ + Enter
    Try one
    OUTPUT / coverage report local · no network

    SEC.08 / QUERY

    Ask, and get a cited answer

    TF-IDF retrieval over 68 facts compiled from profile.json. There is no generation step, so it cannot invent an answer — it either cites one or says it does not have one.

    QUERY / grounded retrieval corpus 68 · sources cited
    Presets

    Ask a question, or press a preset.

    Everything here runs in this tab: TF-IDF over the fact corpus, no network call and no language model. That is a design choice, not a limitation — a retriever that cannot generate also cannot hallucinate, so every answer on this page is a sentence from the profile with its source attached.

    SEC.09 / CONTACT

    Open to industrial AI roles

    Bengaluru first, then Pune, Hyderabad, Mumbai and NCR. Onsite, hybrid or remote. International remote only.