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_e4b() # google/gemma-4-E4B-it, auto-downloads, runs on CPU/MPS/CUDA
# gen = rsa.providers.gemma_e2b() # google/gemma-4-E2B-it, the lighter 5B variant
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.
On the Gemma E4B injection path the prompt is NEUTRAL (no persona, no style words): the human voice comes from the RL latent itself, not prompt engineering.
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.).
Metadata
Release files for rl-sales-augment 0.9.7
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| rl_sales_augment-0.9.7.tar.gz | 68.9 kB | Without provenance | Details for rl_sales_augment-0.9.7.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| rl_sales_augment-0.9.7-py3-none-any.whl | Python 3 | none | any | Details for rl_sales_augment-0.9.7-py3-none-any.whl |
Total release size: 125.7 kB
Release files / rl_sales_augment-0.9.7.tar.gz
| Download URL | rl_sales_augment-0.9.7.tar.gz |
|---|---|
| Size | 68.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
8e8ae35740f649991d5d8d267c79e60c03c36c8a1fca47ad6c75d70287e7f673
|
|
BLAKE2b-256 checksum How to use checksums |
6e16ac444c0a4a2b18b452b0e65e8f578c6319dbbc4c5618d71d1956af06ae83
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.9.6
|
Release files / rl_sales_augment-0.9.7-py3-none-any.whl
| Download URL | rl_sales_augment-0.9.7-py3-none-any.whl |
|---|---|
| Size | 56.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
b96b8a7007f434c7afa7ea3bc93370180afdf360f9112c9e4906fad4b53cadb8
|
|
BLAKE2b-256 checksum How to use checksums |
cfbc22c9a4e5de13be84066266655af7b68153aa391acf02079d4f1752fd3eb0
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.9.6
|
Security for rl-sales-augment 0.9.7
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.