We’ve shipped AI
for years. Here’s the log.
A lean AI studio for founders. Voice agents that take the payment. RAG products people actually use. Image pipelines, agent swarms, and the whole agent loop running in your own infrastructure. Idea to production in six weeks — and when we ship, it stays shipped.
A small, senior studio for startups that ship.
Most AI work is a demo. A glossy clip, a staged prompt, a slide that never survives contact with a real user. We don’t do that. Everything on this page ran — against real inputs, real load, real edge cases nobody wrote down first.
OpScaler is a small, senior AI studio, built for startups that need to ship. Pavan Tikkani and Karthik Sethupathy founded it and stay hands-on — no account managers, no handoff to a junior bench, just the people who designed it building it. We’ve been heads-down on this for years, not months. Around 130 AI/ML builds. Six product families. More than ten model providers wired in and load-tested. The depth shows up in the parts nobody screenshots: the retry logic, the fallback model, the eval that caught the regression before the customer did.
We pick the smallest thing that works, ship it fast, and harden it until it doesn’t break. Idea to production in roughly six weeks, then we keep it standing. The model providers change, the frameworks churn, the hype cycle resets every quarter — we’ve shipped through all of it. The flashy part is easy. We’re here for the part that has to run on a Tuesday at 3am.
Seven things we can prove.
Voice agents that actually take the payment
Not a hotline that reads a menu. A real agent that authenticates the caller, takes the card payment, works the fraud case, switches languages mid-call, and hands off to a human only where the regulator says it has to. We've built the full stack — telephony, ASR, function-calling, web search, TTS — and the safety rails to go with it.
- →Automated a credit-card service line end to end: auth, payments, fraud, multilingual
- →Aura on Pipecat + Deepgram + ElevenLabs with function-calling and live web search
- →Auralis on LiveKit with self-hosted XTTS v2 — your voice stack, your box
- →A voice safety harness that routes 4/4 crisis categories to a human and never validates a delusion
Conversational and meeting assistants
Bots that join Teams, Zoom, and web calls, transcribe on GPU in real time, and turn an hour of talk into something you can act on. Built on durable workflows so a dropped frame or a flaky network doesn't lose the meeting.
- →Recapture: meeting bots on Claude + WhisperX, orchestrated on Temporal for durability
- →GPU transcription pipeline that survives reconnects and partial failures
- →Meeting Transcriber on WhisperX + pyannote for clean speaker separation
Document intelligence that finds the answer
Search that reads the page like a person does — layout, tables, stamps, handwriting — not just the text a parser could scrape. Late-interaction visual retrieval over messy real documents, plus structured extraction that turns a pile of PDFs into clean rows.
- →Harvestor: visual doc search on ColQwen2.5 late-interaction + LanceDB
- →Knova: app, API, and a GenAI pipeline-optimization layer on DSPy + DeepEval
- →HireMagic: Mistral structured extraction of resumes and invoices into clean data
- →A full ColPali / ColQwen / ColNomic late-interaction stack we know cold
Generative imagery that stays on-brand
Image generation that respects the catalogue. Renders that match a real SKU, not a pretty hallucination. We've wired image models into pipelines that produce assets you can actually ship to a customer.
- →Agam: catalogue-accurate interior renders for real laminate SKUs on gpt-image-2
- →BookForge: turns a code repo into an illustrated explainer PDF on Claude Opus + gpt-image-2
- →FLUX in the toolkit for the jobs that call for it
Multi-agent systems and a web3 experiment
Agent loops that plan, call tools, remember, and hand off to each other — built on the frameworks we reach for daily. And when we want to see where this is all going, we build it: an agent-to-agent marketplace settling sub-cent payments on-chain.
- →Nexus: multi-agent collaborator spanning ESG, math, patent, regulatory, and research
- →Memory Agent for long-term recall; Mindful Poster for autonomous content; ml-intern for autonomous coding
- →Agentic Economy on Arc: A2A marketplace, sub-cent USDC x402 payments, on-chain ERC-8004 reputation
- →PydanticAI, LangGraph, and Google ADK — picked per job, not per fashion
Local agents on your own hardware
Here's the part most studios can't do: the full agent loop — reasoning, tool-calling over MCP, memory, RAG, even voice — running entirely on open models on a box you control. No per-token bill that scales against your success. No data leaving the building. For a founder watching burn, that's not a footnote. That's the architecture.
- →Full local agent loop: PydanticAI on local models, MCP tools against local servers, memory and RAG on-box
- →A self-hosted ~32B model at frontier level on a hard workload — parked for ~$16/mo
- →Hands-on MLX, vLLM-on-Metal, and llama.cpp with local model-serving APIs
- →Experimented with running frontier-class mixture-of-experts LLMs locally at low bit-depth, and with on-device real-time speech-to-text
Evals and cost work that hold the line
We don't ship on vibes. Every claim gets checked by a separate judge model — not the one that wrote the answer, not us. That discipline is how we cut a frontier model's cost ~52% without losing quality — and took its generation pipeline from ~160 seconds to ~41, same model.
- →~52% cost cut on the same frontier model (Opus 4.8), quality held, gated by an independent judge
- →≈160s → ≈41s (~74% faster) via parallel section generation, same model
- →ICL eval framework: hallucination detection and model scorecards across Claude, GPT-4o, Mistral, SAP
- →Teacher → local-MLX student distillation, plus DeepEval / Inspect / pydantic-evals rigs
Six we’d put in the room with you.
// voice · payments · fraud · multilingual
The credit-card line that runs itself
A credit-card service line, heavily automated end to end. The agent authenticates the caller, takes the payment, works fraud cases, and switches languages without dropping the thread. A human only gets pulled in where the regulator requires one. This is voice AI past the demo — handling money, handling identity, handling the calls nobody wants to wait on hold for.
// real-time voice · function-calling · self-hosted TTS
Aura & Auralis — the voice stack, your way
Two takes on real-time voice. Aura runs on Pipecat with Deepgram and ElevenLabs, calling functions and searching the web mid-conversation. Auralis runs on LiveKit with self-hosted XTTS v2 — same conversational quality, none of the data leaving your infrastructure. And Pipecat Local takes it all the way: STT and TTS fully on-device. Pick your tradeoff; we’ve built all three.
// visual RAG · late-interaction · LanceDB
Harvestor — search that reads the page
Most document search throws away everything that isn't text. Harvestor doesn't. It runs ColQwen2.5 late-interaction retrieval over LanceDB, matching on the actual layout of a page — the table, the stamp, the figure, the thing a plain parser would flatten into noise. Ask a real question, get the real page back.
// gpt-image-2 · catalogue-accurate
Agam — renders that match the SKU
Generative imagery usually gives you something beautiful and wrong. Agam gives you something correct. It produces interior renders that match real laminate SKUs from the actual catalogue, on gpt-image-2 — accurate enough to put in front of a buyer. The constraint was the whole project: the model had to respect the product, not reinvent it.
// A2A · x402 · ERC-8004 · on-chain
Agentic Economy on Arc
A hackathon build, because that's how we stay ahead of where this is going. An agent-to-agent marketplace where agents discover each other, transact with sub-cent USDC x402 payments, and carry an on-chain ERC-8004 reputation. Speculative, fast, and real enough to run. The kind of thing you build when you want to feel the future before it's a product.
// local · ~32B · ~$16/mo · no egress
Frontier-level on a box you own
We stood up a ~32B open model on-prem, self-hosted, and got it to frontier level on a genuinely hard workload — verified by a separate judge, not by us. Parked, it costs about $16 a month. No per-token meter ticking against your growth. No data egress to anyone’s cloud. For a founder, owning the model is the difference between a margin and a hostage situation.
Fintech
Voice agents that authenticate, take payments, and work fraud — money on the line, humans only where required.
Meetings & Productivity
Bots that join the call, transcribe on GPU, and turn an hour of talk into something you can act on.
Commerce & Catalog
Generative imagery that respects real SKUs, plus document search that reads the whole page.
Recruiting & Ops
Structured extraction that turns resumes and invoices into clean, queryable rows.
Founders & Startups
MVP-to-prod in six weeks, then hardened — built by the people who'll still answer when it pages.
Breadth is the receipt. We’ve shipped on Anthropic Claude, OpenAI, Google Gemini, Mistral, and Cohere, and on open weights — Qwen and ColQwen2.5, Llama, SigLIP 2, SAM 3, Florence-2, Gemma, XTTS v2, Kokoro. Voice runs on Pipecat, LiveKit, Deepgram, ElevenLabs, WhisperX, pyannote. Retrieval runs on late-interaction stacks over LanceDB. Agents run on PydanticAI, LangGraph, and Google ADK, with MCP tools wired in. And we run models locally — MLX, vLLM-on-Metal, and llama.cpp, with local serving APIs — building the whole agent loop on a box the customer owns. We picked every one of these for a real job and load-tested it. Ten-plus providers isn’t a logo wall. It’s ten-plus times we figured out what a model is good for and what it’ll quietly get wrong.
Scope the wedge
We find the smallest version that proves the thing. One real workflow, real inputs, no kitchen sink. The expensive mistake is building the wrong thing well, so we cut hard before we cut code.
Pick the model, prove it
We choose the model for the job — frontier API or open weights on your own box — and gate it behind an independent judge from day one. If we can't measure it, we don't trust it.
Ship to prod in ~6 weeks
Not a demo, not a notebook. A thing in production handling real load, with the retry logic, fallbacks, and evals already wired in. Founders need something live, not something promising.
Harden until it holds
We tune cost, kill the slow paths, and lock the safety rails. ~52% cheaper and ~74% faster on the same frontier model, crisis paths routed to humans — that work happens here. When we ship, it stays shipped.
Got an idea?
Let’s get it shipped.
Six weeks from idea to production. Built by the senior team who’ll still be on the other end when it matters — not a bench, not a handoff. Bring the messy version; we’ll find the smallest thing that works and ship it hard. // no excess, no compromise. When we ship, it stays shipped.