AI consulting · agentic deployment · SLM tuning · AI memory

Train big once.
Deploy small forever.

AI that learns your business: small models tuned to your work, with memory that compounds. Move off pay-per-token frontier APIs onto software you own — same-or-better results at a fraction of the cost, every claim backed by published benchmarks.

Token costs ↓
tuned SLMs · ~$0/token
OPEX ↓
agents do the repeat work, once
Revenue ↑
software that learns your business
SEQ · Memory layer activeLive
Hey — picking up where we left off on the integration?
Session resumed · 12 memories injected● Persistent
Linux native
0 ms
cold-start memory

What we work on daily

High-volume, per-item work — where a tuned small model pays for itself.

01

Document & record extraction

Legal · medical · financial

02

Customer support automation

Contact-center, tier-1 deflection

03

Transcription → summarization

Scribe · meetings · calls

04

Batch classification

Moderation & routing

05

Content generation at scale

SEO · product · e-comm

06

Compliance & security scanning

Per-line review

07

Outreach personalization

SDR / leads

08

Translation & localization

Any language pair

Agentic
by default
Open source
one product
0M+
lines shipped
Linux
native

The research lane

We're training AI
to manage its own memory.

Most agents forget. Ours don't. We publish what we learn building persistent intelligence — the methodology, the datasets, the benchmarks against frontier models. The same architecture goes into every system we ship.

In progress
Autonomous memory compression

Training a small Linux-native model to classify, compress, and store conversational memory in real time — replacing frontier API calls in the persistence layer of every agent we build.

Target: match Opus 4.7 accuracy at 1/200th the cost. Open eval coming.
In progress
Self-directed fine-tuning

An agent that uses its own accumulated decisions as training data, scaffolds eval pairs from its own corrections, and trains its next body without a human in the labeling loop.

Live training log + eval dashboard publishing here as benchmarks land.
Read the research →Public dashboard launching as the first benchmarks land.

Case study — measured, not quoted

Measure your model.
Don't trust the leaderboard.

We ran two leading ≤4B open models — Qwen3-4B and Gemma-4 E4B — head-to-head as a local agent brain on an Apple M4 Mac mini (16 GB), Q4 via ollama. Not leaderboard trivia: an objective, machine-scored harness (grep / parse, 3 trials each) measuring what actually breaks agents in production — routing, clean tool-call JSON, and complete hand-offs.

Tool-call reliability
0/9
Qwen3-4B — clean tool calls
9/9
Gemma-4 E4B — clean tool calls
Checks passed per task (out of 3 trials)Qwen3-4BGemma-4 E4B
3
2
1
0
3
3
0
3
0
3
0
3
3
3
Routing
JSON pure
JSON valid
JSON intent
Draft prompt

The three JSON / tool-call tasks are the cliff: Qwen 0/3 on every one, Gemma 3/3 on all fifteen.

Overall score (15 checks total)
Gemma-4 E4B15/15
Qwen3-4B6/15
Median speed (tokens / sec) 24 = usable
32.7
26.5
Qwen3-4B
Gemma-4 E4B

Speed didn't decide it — both cleared 24 tok/s, and Qwen was actually faster. Reliability decided it.

The finding

Qwen3-4B — even with thinking mode off — refused to emit a clean tool call. It monologues its reasoning in the response body (“We are given a query… Step 1…”) and never produces parseable JSON. Gemma-4 E4B followed the tool-call contract every time (15/15), at comparable speed.

The thesis

A “smarter” model that won't emit a parseable tool call is useless for an agent. Model selection must be measured on your production criteria — objective, reproducible, machine-checked — not brand or leaderboard. We even caught our own bias: we leaned Qwen; the measurement flipped the decision. Proof over preference is what we sell.

Measured 2026-07-06 on an Apple M4 Mac mini· every number is re-runnable · benchmark reproducible on request.

Fine-tuning capability

The multiplier,
and the minefield.

A properly fine-tuned small model is a force multiplier. Done wrong, it looks like it worked and doesn't. We treat “it works” as a machine-proven claim — never a vibe.

75%95%

A Qwen3-4B classifier we fine-tuned in-house — 75% → 95% on its task, in a single short run. A 4B model can match or beat a model 30× its size on your narrow domain, running local, offline, and cheap — on hardware as modest as a 6 GB-VRAM laptop. You own the weights and the data.

The benefit
  • Beats models 30× largeron a narrow, well-scoped domain.
  • Local, offline, low-latencyruns on a 6 GB-VRAM laptop — no API bill.
  • You own itthe weights and the training data are yours.
The danger — why most fine-tuning quietly fails
  • Catastrophic forgettingIt gains your task and loses general ability.
  • OverfittingMemorizes a tiny dataset, fails on anything new.
  • Tool-call / format regressionThe exact failure above — induced by bad tuning.
  • Evaluation on training dataThe classic false-green: “97%” on data it already saw.
  • Data leakageSecrets or PII baked into weights you can’t un-bake.
Why Riscent is good at it
Base chosen by measurement

We pick the base model the way we ran the study above — not by brand.

Held-out, red-first gate

The tuned model must pass objective checks it never trained on before we ship it.

Attested, not assumed

Every claim maps to a re-runnable test — coverage and verification are proven.

Local by default

Your proprietary data never leaves your hardware.

The one honest trade: fine-tuning is powerful because it's narrow. We scope it, prove it on held-out data, and tell you plainly where the model's competence ends. That honesty is the product.

The deployment model · real-world SLM economics

Train big once.
Deploy small forever.

Renting intelligence by the token is a bill that grows with every customer you serve. Our model flips it: use a big model once — to set the bar and generate the training data — then run a small model you own, local and offline, at near-zero cost per call. Lower operating cost, higher close rate.

Train big
once

A frontier model earns its keep one time: it defines the task, generates and curates the training data, and sets the quality bar. You pay the big model to teach — not to run in production.

Deploy small
forever

A fine-tuned small model hits that bar on your narrow job, then runs local, offline, and fast — at near-zero cost per call. You own the weights. The recurring bill stops, and it keeps working when the internet does not.

75 → 95%

accuracy — a 4B we fine-tuned, in one short run

30× smaller

a tuned 4B matching models 30× its size

15 / 15

tool-call reliability that actually completes the action

$0 / token

no API bill once it runs on hardware you own

Lower OpEx

The per-token invoice that grows with your traffic becomes a one-time training cost plus cheap, owned inference. No frontier bill scaling with every conversation. Runs on hardware as modest as a 6 GB laptop — even offline.

Increase revenue

A fast, reliable, always-on model completes the actions that make money — answers the call, books the appointment, captures the lead — instead of stalling on a broken tool call at 3 a.m. Reliability is revenue.

Own your model. Lower the bill. Close more.

Two ways to move: have us build and ship it, or bring us in to advise. Either way, it's measured — we show you the numbers before you commit.

See the full case for SLM →

The build lane

Want this engine
working on your problem?

We take on a small number of build engagements every quarter. Same team that's publishing the research. Same architecture, applied to whatever you're shipping.

Custom software
Full product builds

Agentic backends, structured-data flows, internal tools, vertical SaaS. We architect, we ship, we hand over what works.

Web apps
Modern web platforms

Next.js, Postgres, auth, payments, dashboards. Production-grade from day one — the stack we use for our own products.

PWA integration
Install-anywhere apps

Progressive Web Apps that ship to iOS, Android, and desktop from one codebase. Offline-first when it matters.

Book a call →We'll show you something working before you sign anything.

The problem

Sound familiar?

Your AI feature is a wrapper around someone else's API. So is your competitor's.

We architect the persistence layer underneath the wrapper. Memory is the moat — and it compounds.

Every session your agent forgets the user. Every conversation starts cold.

We build the memory pipeline: extraction, embedding, recall, decay. Your agent gets better every week.

You're burning frontier-model tokens on tasks a small model could do.

We fine-tune Linux-native models on your data, swap them in behind the same interface. Same quality, 1/200th the cost.

Your team can ship a feature. You need someone who can ship an agentic system.

Architecture, training, infra, evals, production. One team. We publish the methodology so you can audit the work.

Results over idealism

Most AI dies in the pilot.
We build what ships.

The national numbers are sobering — and public. Most AI spend never reaches the P&L. That's the idealism gap: pilots, demos, and AI theater that never touch the bottom line.

95%

of enterprise generative-AI pilots deliver no measurable P&L impact — only ~5% break through.

MIT NANDA · “The GenAI Divide,” 2025
39%

of organizations see any enterprise-level earnings impact from AI — and typically under 5%.

McKinsey · “The State of AI,” 2025 (≈2,000 firms, 105 nations)
Back-office

is where the real ROI sits — automation and engineering, not the demo the budget chased.

MIT NANDA, 2025

We build for the 5%. Our judgment is done by experience, not fresh-graduate idealism — wisdom that has grown past “AI will fix it” and is built on the foundation of shipped results. Here's the proof, not the pitch.

Shipped · open source
Phantom Vault

An open-source secret vault for the age of AI agents: the model uses your API keys by name and never sees the value — encrypted at rest, jailed at runtime, sanitized on the way out. Free, Apache 2.0.

phantomvault.riscent.com
In every system we build
Synthetic memory

The persistence layer that makes an agent remember: extraction, embedding, recall, decay. Agents stop starting cold and get better every week. Memory is the moat — and it compounds.

See the research
And the models are chosen the same way — by measurement
15 / 15

Gemma-4 E4B on our tool-call harness — Qwen3-4B managed 6/15

75 → 95%

a 4B classifier we fine-tuned, in a single short run

30× smaller

a tuned 4B matching models 30× its size — local and offline

See the full benchmark →
Six products shipped on the same stack:
ChatterboxBookBotLinguaReachVoiceGuardDripForceVoiceTrain

Healthcare  ·  Home Services  ·  Legal  ·  Real Estate  ·  Restaurants  ·  Auto  ·  Fitness  ·  E-commerce

AI visibility — our specialty

Your buyers stopped Googling.
They're asking the models.

People don't scroll ten blue links anymore — they ask Gemini, Claude, or ChatGPT and act on the answer. Here's the catch: an AI can only recommend what it was trained on. If the big three models don't know your business, you're not in the running — you're invisible. And that gap is closing fast.

−25%

of traditional search volume, gone by 2026, as buyers move to AI chatbots

Gartner, 2024
1.6B

ChatGPT queries a day — already ~12% of Google’s search volume

Industry reporting, 2025
58.5%

of U.S. searches now end with zero clicks — the answer never leaves the page

Similarweb, 2025
1 in 3

Gen Z (and 1 in 4 millennials) go to AI over other channels for shopping advice

Commerce / Future Commerce, 2025
47%

of Gen Z discovered a new brand through ChatGPT this year

PR News, 2025
+4,700%

growth in shopping-related AI searches, July 2024 → July 2025

Adobe / industry data, 2025
How the new front page works

Ask an AI “who's the best [your category] near me?” and it names a handful of businesses — chosen from what it was trained on. There is no page two. Being absent isn't neutral; it's being left out of the only answer the buyer sees.

Soon that window narrows further as the models consolidate their sources. Getting into that training data — before your competitors do — is the specialty. It's the front and forefront of where discovery is going.

Gen Z product research is already a coin flip
Share who use each channel to research a product
Search engines37%
AI (ChatGPT, Gemini…)33%

Nearly even — and AI-driven shopping searches grew +4,700% in a year. The crossover isn't a forecast; it's arriving. Source ↗

The Riscent AI Visibility Scan

Find out what Gemini, Claude, and ChatGPT say about you.

We test what each of the big three models actually says about your business and your category, show you exactly where the gaps are — and then get you into the answers. This is what we do: the same team that measures models and ships fine-tuned ones, pointed at your visibility.

Scan my business →
Proof of concept

Test it yourself. Ask ChatGPT or Gemini “Tell me about Ryan Bolden and InboundAI365.” The models answer — because we engineered that knowledge into the places they train on. We do exactly the same for your business.

Why us

We build the kind of AI
most teams don't know how to build.

Most AI work today is wrapper code around someone else's API. Ours isn't. We architect persistence, train our own models, publish the methodology, and run it in production. Healthcare was our first vertical — HIPAA-grade voice agents, midnight calls, two languages — and the same architecture now ships across eight industries.

2M+
lines shipped
Agentic
by default
Open source
one product
HIPAA
production-grade

Healthcare  ·  Home Services  ·  Legal  ·  Real Estate  ·  Restaurants  ·  Auto  ·  Fitness  ·  E-commerce

Two ways
to build with us.

We're a consulting agency — no products to buy. Custom builds: we architect and ship your agentic systems, fine-tuned models, and AI software. Consultations: we advise on model selection, memory, and getting your business visible inside the models buyers now ask. Either way, you work with the team publishing the research.

Limited engagements each quarter · Response in 24 hours