01 · Claude API consultancy
Get Claude into production — properly.
You've prototyped with the API. The POC works. Now you need it to handle real traffic, real cost, real evals — and not embarrass you on a Friday. We've shipped Claude into four production stacks and we know where the rough edges are.
- — System-prompt architecture, cache breakpoints, tool-use design
- — Prompt evals (golden sets, regression suites, cost per turn)
- — Cost guardrails, retries, fallback chains, structured output
- — Migration between Claude model versions without prompt drift
- — Embedded with your engineers — not a deck handoff
Who it's for: Engineering teams with a working Claude POC who need a senior pair to ship it.
Example outcome: A coaching network's WhatsApp agent — cache hit rate from 0 to 72%, turn cost down 4x, zero late-night replies missed.
We share pricing on the diagnose call. No retainer minimums, no SOW theatre.
02 · AI agent development
Agents that ship.
Most "AI agents" are a demo on someone's laptop. Ours run 24/7, talk to real customers, and have evals you can show a board. We build the boring scaffolding — tool registries, memory, escalation, observability — so the model gets to do the smart part.
- — Surface: WhatsApp, Slack, CLI, web, voice (we pick what fits)
- — Tool-use, memory, handoff-to-human, fallback behaviour
- — Evals before launch and continuous after — drift is real
- — Per-turn cost tracking, rate limits, escalation rules
- — DPDP/GDPR-clean from day one — consent, retention, right-of-access
Who it's for: Teams who need an agent that actually runs in production — not a Friday-afternoon Notion demo.
Example outcome: Chronos Auto — multi-tenant WhatsApp agent serving four client businesses on one backend, founder-grade tone per tenant.
We share pricing on the diagnose call. Fixed scope, no per-message surcharges.
03 · Custom MVP (14-day) · The headline
Working software in 14 days.
This is what we're known for. You bring the problem, we build the heart of the software, and at day 14 you have something working in your hands with your real data. If it's not right, you pay nothing — not for the call, not for the build, nothing.
- — Day 0: 30-min discovery. We say yes or no honestly.
- — Day 1: 1-page scope. What's in, what's deliberately out.
- — Day 2–14: We build. You stay close. Real data, real users.
- — Day 14: You decide. Free walk-away, or 50/50 to finish it.
- — Web app, mobile app, agent, desktop tool, automation — whatever shape fits
Who it's for: Founders and operators with a specific problem no off-the-shelf tool solves.
Example outcome: Inventory CRM for a power-electronics manufacturer — built in 14 days, cut stock-out incidents 64% in the first quarter.
We share pricing on the discovery call. No surprise add-ons, no upfront-and-pray.
04 · WhatsApp automation
WhatsApp that talks like your best closer.
Inbound at 11 PM, qualified before breakfast. Mid-tier WhatsApp automation that doesn't read like a chatbot, doesn't ghost on edge cases, and hands off to a human at exactly the right moment. This is what Chronos Auto does — we run our own product on the same engine.
- — Lead qualification with founder-trained tone — not generic AI voice
- — Support, broadcast, opt-in handling, 24h-window discipline
- — Per-lead memory, follow-up sequencing, escalation to the right person
- — Multilingual / mixed-language detection — replies match the inbound
- — DPDP-clean consent flow, opt-out, 30-day retention purge built in
Who it's for: Teams losing leads to slow response — real estate, clinics, service businesses, D2C, agencies, SMBs.
Example outcome: A coaching network — 1,400 leads qualified in the first 6 weeks, zero late-night replies missed.
We share pricing on the diagnose call. No per-conversation surcharges, no seat audits.
05 · Tax-notice automation
Notices, sorted before the deadline.
Desktop tooling for CAs and tax practices drowning in IT, GST, and TDS notices. Local-first — the documents never leave the partner's machine. OCR-driven intake, deadline tracking, client-aware drafting. This is what Chronos CA does — we built it because our own CA kept losing Saturdays to it.
- — OCR ingest of IT / GST / TDS PDFs, with structured field extraction
- — Deadline calendar, escalation alerts, partner-level dashboards
- — Client folder, history, draft-reply scaffolds tuned to your firm's tone
- — Local-first desktop — documents stay on the partner's machine
- — DPDP-clean by design — no third-party cloud OCR
Who it's for: CA firms, tax practices, in-house tax teams handling 50+ notices a season.
Example outcome: A 5-partner CA firm — 230 GST notices handled in 14 days, full exposure surfaced, all responded within deadline.
We share pricing on the diagnose call. No per-notice fees, no per-seat audits.
06 · Agentic memory systems
Memory that recalls.
RAG that works isn't a vector store with a re-ranker bolted on. We design production memory: hybrid retrieval, parent_id dedupe, graph relations, live decay, lens-prefixed reranking. The kind of architecture that holds up when the corpus crosses a thousand documents and the agent has to remember what mattered yesterday.
- — Hybrid retrieval (dense + sparse + graph), 13-stage pipeline
- — parent_id dedupe — no more duplicate chunks polluting context
- — Lens-prefixed reranking + 6-term live decay for involuntary recall
- — Sub-50ms recall on indexed corpora up to ~10k documents
- — Evals: top-k hit rate, MRR, recall@3 — measured, not guessed
Who it's for: Teams whose agent loses the plot at document 500. RAG that needs to be production, not a workshop demo.
Example outcome: File-finder across 1,547 personal documents — 93/95 top-3 hit rate, sub-50ms involuntary recall. See the architecture →
We share pricing on the diagnose call. Architecture review or full build — both shapes available.