rl-sales-augment

A trained RL policy that picks the next sales move (rapport / pitch / objection / discount / close); your LLM writes the words. The trained model downloads once on first use (sha-verified, cached): CPU, no finetuning, works with any LLM.

pip install rl-sales-augment

Quickstart (local Gemma, no API key)

The policy was trained alongside Gemma 4; running it fully local (E2B or E4B) is the first-class path:

pip install "rl-sales-augment[gemma]"

import rl_sales_augment as rsa

gen = rsa.providers.gemma_e2b()      # google/gemma-4-E2B-it, auto-downloads, runs on CPU/MPS/CUDA
# gen = rsa.providers.gemma_e4b()    # google/gemma-4-E4B-it, the bridge-aligned flagship
bot = rsa.load_agent(gen, company_ctx="Acme sells AcmeBox, an $8k on-prem appliance.")

out = bot.reply("honestly it feels expensive vs AWS")
out["chosen_move"]   # 'RAPPORT'  <- the RL decision
out["belief"]        # {'interest': .5, 'trust': .5, 'budget_fit': .2, 'objection': .8, ...}
out["reply"]         # the LLM's words, executing that move

Prefer an API model? Same code, different one-liner:

gen = rsa.providers.openai_chat(model="gpt-5.5")      # or gemini_api() / anthropic_chat()

bot.reply() is stateful: keep calling it, the agent remembers. For stateless use (e.g. behind an API), pass the whole conversation in OpenAI message format:

out = bot.chat([
    {"role": "user", "content": "what does it cost?"},
    {"role": "assistant", "content": "Depends on seats. How many do you need?"},
    {"role": "user", "content": "40 seats, but budget is tight"},
])

Providers

rsa.providers.gemma_e2b()                                     # [gemma]   local Gemma 4 E2B, no API key
rsa.providers.gemma_e4b()                                     # [gemma]   local Gemma 4 E4B, no API key
rsa.providers.openai_chat(model="gpt-5.5")                    # [openai]  OPENAI_API_KEY
rsa.providers.anthropic_chat(model="claude-sonnet-5")         # [anthropic] ANTHROPIC_API_KEY
rsa.providers.gemini_api()                                    # [gemini]  GEMINI_API_KEY
rsa.providers.gemini_vertex()                                 # [gemini]  gcloud ADC + GCP_PROJECT
rsa.providers.openai_chat(base_url="http://...")              # any OpenAI-compatible server

Or bring your own: any gen(prompt) -> str works.

Multilingual: the bot replies in the customer's language automatically (tested: Malayalam, Hindi, Tamil, Japanese, Spanish, German; a built-in script detector keeps even small local models on-language for Indic/CJK/Arabic/Cyrillic scripts). Romanized Indic is supported too: Manglish / Hinglish / Tanglish ("ntha visesham, sugano?") is detected and mirrored on frontier models.

API keys (.env)

Put credentials in a .env file next to where you run your script; providers load it automatically (real environment variables take precedence). Never commit it.

# .env
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GEMINI_API_KEY=AIza...
GCP_PROJECT=my-gcp-project        # for gemini_vertex (gcloud ADC)

Explicit control: rsa.load_env("/path/to/.env"). Local Gemma needs no key at all.

MCP server

pip install "rl-sales-augment[mcp]"
rl-sales-augment-mcp        # stdio

{ "mcpServers": { "rl-sales-augment": { "command": "rl-sales-augment-mcp" } } }

Tools: next_move (buyer state → RL move), perception_prompt, list_moves, list_segments.

REST API

pip install "rl-sales-augment[gemini,api]"

A complete FastAPI server (POST /v1/chat, OpenAI-format messages) ships in examples/fastapi_server.py.

Why not just call GPT-5.6 / Opus 4.8 / Gemini directly?

Because what kills LLM sales conversations isn't the words, it's the timing. Frontier models are trained to be helpful and agreeable, so on a skeptical buyer they answer every objection politely, forever, and never risk asking for the deal (measured: 0/4 closes on adversarial buyers while handling every question beautifully). A bigger model writes better sentences; it doesn't fix this, because next-token training never rewards a deal that closes six turns later.

The bundled policy is different in kind, not degree:

• Trained on outcomes, not text. PPO over millions of simulated deals with delayed, stochastic rewards. It has lost deals to premature pitching, burned reputation on spam-closing, and learned that discounting converts SMBs but insults enterprise. An API model has read about selling; the policy has sold.
• State-dependent timing. Prompting "be assertive, always close" makes a bot uniformly pushy. The skill is when: the policy closes at high readiness and keeps building trust below it. Same LLM writing the words, right moment to ask. Result: 100% vs 19-31% close in a paired A/B, 3/4 vs 0/4 on hard buyers (simulated; harness in the repo).
• Consistent and auditable. Sampled LLM strategy swings run-to-run; the policy is deterministic, and every turn logs chosen_move + belief, so you can see why it did what it did.
• Complementary and tiny. A ~1MB MLP on CPU. Keep GPT-5.6 / Opus 4.8 / Gemini for language, empathy, and knowledge; add the decision layer they don't have. Retrainable on your own funnel's economics (the commercial offering).

Details, transcripts, and a 67-second demo: github.com/NandhaKishorM/rl-sales-augment

License

AGPL-3.0-or-later. Training the policy on your own market is the commercial offering: nandakishor@convaiinnovations.com (Convai Innovations Pvt. Ltd.).